What's New in Local AI

Alibaba split its Qwen3.8 generation into two open releases this month, and they make very different promises. Qwen3.8-27B, published August 14, is a dense 27B model with a native 262K-token context, multimodal on image and video, under a plain Apache 2.0 license, no thresholds, no rider, yours to self-host and ship. Two days earlier, Qwen3.8-2.4T-A95B brought Qwen3.8-Max's flagship weights to Hugging Face: a Mixture-of-Experts at 2.4 trillion total parameters and roughly 95 billion active, text-only, without the API version's 1M-token context. Alibaba published its weights under a new "Qwen3.8-Max License" instead: prominent attribution once a product clears 100 million monthly users or $20 million in monthly revenue, and a separate paid license if you resell it as a hosted service and clear $50 million in revenue over any 12 months. Both are labeled open. Only one of them actually behaves that way. The 27B is the one worth building on if you want to own what you ship; the 2.4T is worth reading closely before you do.

The runtime side kept catching up to the last two weeks of open model releases. vLLM 0.27.0 (August 10) shipped a full Kimi K3 stack in one release: model files and kernels, Python and Rust frontends, quantized-checkpoint support, and a way to shard the shared expert instead of replicating it, alongside a stack of DeepSeek-V4 performance work covering sequence parallelism, kernel launch savings, and time-to-first-token cuts across several code paths. 0.27.1 followed a day later with a small quantization fix. llama.cpp's automated build b10427 (August 14) fused the gate and up matrix multiplies with the GLU activation for Intel Arc GPUs on the SYCL backend, measured at roughly 2 to 12 percent faster generation depending on batch size. None of this is a new model. It is the runtime layer doing the unglamorous work of making sure a model you can already download runs faster on hardware you already own, which is the actual bottleneck for most local setups, not the weights.

Cline kept shipping through the week, three point releases (4.1.8 on August 11, 4.1.9 on August 13, 4.1.10 on August 14) aimed mostly at reliability rather than features. The headline fix: two Cline installs on different builds could shut down each other's Hub daemon in a loop, killing live sessions with an abnormal socket close; build identity is now compared through a total order so at most one side can win, and a Hub still serving a session is left alone until it goes idle. Cache-heavy sessions had been reporting token totals inflated roughly 5x because the counter re-counted the whole cached conversation on every request, now fixed to report disjoint per-request buckets. Vertex users get Fable 5 support and the full per-region model catalog instead of a hardcoded list. And 4.1.10 adds an opt-in toggle to let supported models search the web mid-task, with results persisted in the conversation. None of this is glamorous, but a Hub daemon that kills its own sibling process is exactly the kind of bug that makes a self-hosted setup feel fragile, and BYOLLM tools live or die on that trust.

Last week's warning had no numbers attached. This week DeepSeek's own pricing page filled them in. Starting August 16 at 16:00 UTC, the API moves to peak and off-peak billing, with peak hours running 01:00 to 04:00 and 06:00 to 10:00 UTC and off-peak set at half the peak rate. V4-Flash output climbs from a flat $0.28 per million tokens to $0.66 off-peak and $1.32 at peak, roughly 2.4x and 4.7x the current rate. V4-Pro follows the same shape, output rising from $0.87 to $1.98 off-peak and $3.96 at peak. Input and cache pricing move too, on the same schedule. None of this makes DeepSeek expensive by frontier standards, even $1.32 per million output tokens undercuts most closed labs, but the shape of the story hasn't changed from last week: a price you were told to plan around is not a price, it is a suggestion the vendor can revise on notice. DeepSeek-V4-Flash is still MIT and still downloadable, so the peak-hour clock only applies if you choose to keep renting.

DeepSeek's other release this month is the one that doesn't make headlines: DeepSeek-V4-Pro-0813 became the official, non-preview release of its flagship model, a Mixture-of-Experts running roughly 1.7 trillion total parameters, a million-token context window, and the same MIT license V4-Flash shipped under in July. MIT means no usage strings, no revenue rider, yours to pull and serve. It also means a serious download, the weights run to hundreds of gigabytes even after quantization, so this is rack hardware, not laptop hardware. That is the honest tradeoff at the top of the open-weight stack right now: the most capable models you can fully own are also the ones few people can actually host. If you have the hardware, V4-Pro is a flagship you never have to send a repricing notice to yourself. If you don't, V4-Flash from three weeks ago is still the one that fits.

Last week's LFM2.5-2.6B was text-only. On August 12 Liquid AI filled in the vision half with LFM2.5-VL-3B, a 3.1B-parameter vision-language model built on the same LFM2.5 hybrid architecture, tuned for screen understanding, grounding, function calling, and multi-image input, and pitched as competitive against vision models roughly twice its size. It ships with day-one support across five runtimes, llama.cpp, MLX, vLLM, SGLang, and ONNX, so it drops into whatever you already run rather than requiring a new one. It is a non-reasoning model that answers directly, which keeps latency low, the tradeoff you'd expect for something meant to run on a phone or a laptop rather than a server. Small, fast, and locally runnable is a genuinely good direction. Liquid pitches it as open-weight and deployable without restrictions, but as with any model you plan to build a commercial product on, read the actual license terms on the model card before you ship.

Two frontier labs moved on price the same week, in opposite directions, and neither move was made by you. Anthropic quietly confirmed that Claude Sonnet 5's launch pricing of $2 per million input tokens and $10 per million output tokens, originally billed as introductory through August 31, is now the permanent rate; a previously scheduled increase to $3 and $15 on September 1 will not happen. On August 12, xAI launched Grok 4.6, built for long-running agent tasks that stay on a problem across many steps, priced at $2 per million input tokens and $6 per million output, with a faster variant at twice that rate. Cheaper is good news either way, and a cancelled hike is better than a hike. But both numbers were set unilaterally, by vendors who can just as easily set them again next quarter. That is the whole difference between a rented model and an owned one: the price of a model on your own machine is not a decision someone else gets to revisit.

OpenAI brought its Daybreak cybersecurity models to Amazon Bedrock on August 11: Daybreak Red, which provides access to GPT-5.6 Cyber, and Daybreak Blue, a safeguarded version of GPT-5.6 Sol. Both are built for vulnerability discovery and exploit reproduction work, and both are locked behind more than a login. Amazon's own announcement says getting access requires enrollment in Daybreak through OpenAI, plus coordination with your AWS account team, on top of the usual legal attestations that gate offensive-security tooling. Amazon does note that no one, including OpenAI, can see customer prompts or outputs, and none of it is used for training, which is a real privacy commitment as these things go. But the access model is the story: the more capable and more sensitive the rented model gets, the more paperwork stands between you and using it. A model on your own hardware doesn't require enrollment approval to run a scan on your own code.

Meta Superintelligence Labs released Muse Glimmer on August 10, its first open-weight model and a real departure from the Llama license's usage caps: this one ships under a plain Apache 2.0 license, no revenue threshold, no acceptable-use rider. It is a 30B-parameter multimodal model purpose-built for local agent workloads, coding agents, and long-running assistants, small enough to run on a single consumer GPU or a Mac. At full precision it needs over 55GB of memory, but 4-bit quantization brings that under 20GB, comfortable inside a 24 or 32GB card, and Meta tested it on an M4 Max, an M5 Max, and an RTX 5090. Ollama shipped initial support in 0.32.7 (August 10) through its MLX engine on Apple Silicon, with NVIDIA, AMD, and broader platform support promised in the coming days. A 30B model that actually behaves like open weights, rather than open-with-conditions, from the lab that has spent the last two years defining what usage restrictions could get away with, is worth noticing on its own. Whether Meta keeps releasing this way is the actual question.

On August 6 DeepSeek told developers to expect a "significant" increase in its API prices, without naming a number or a date, and said to plan usage accordingly; the notice sits on its own API pricing page. It lands just after DeepSeek introduced peak-hour surge pricing, so the cost of using the API has now moved twice in short order. The cause is its own success: reporting around the announcement describes V4-Flash usage growing explosive enough to overwhelm DeepSeek's compute, so pricing below cost stopped being sustainable. Even a tenfold rise on the current $0.14 per million input rate would still undercut the frontier labs, which is probably true, but it is also the whole point. The lab that won the market on price is now telling you the price was never real. A model you run on your own hardware does not send a repricing notice, and DeepSeek-V4-Flash is MIT-licensed and downloadable, so this is one cloud bill you can opt out of entirely.

Two moves from the rented-model side this week, both on August 6. GitHub Copilot added Kimi K3 to its picker across Pro, Pro+, Max, Business, and Enterprise, hosted on Fireworks AI and metered at $3 per million input tokens, $15 per million output, and $0.30 per million cached. Kimi K3 is open-weight, the same 2.8-trillion-parameter checkpoint Moonshot published in July, so this is a model you could pull and serve yourself being sold back to you by the token. Separately, OpenAI made GPT-5.6 Luna the default for free and Go-tier ChatGPT users, with unlimited text chats. Unlimited is a good deal until you remember it still means their servers, their logs, their rate limits, and their definition of unlimited tomorrow. Neither is a bad product; both are reminders that a model you rent is one someone else prices, meters, and can change under you.

Last week DeepSeek's MIT-licensed V4-Flash landed and the local stack rushed to support it; this week that support kept maturing. ExLlamaV3 1.4.0 (August 6) added DeepSeek-V4 support for NVIDIA cards, improved AVX512 CPU-offload performance so a large Mixture-of-Experts can spill onto system RAM, and sped up prefill across all models. llama.cpp's SYCL backend gained DeepSeek-V4 operators for Intel Arc GPUs, extending the acceleration past NVIDIA and Apple. And Ollama 0.32.6 (August 4) realigned its streaming output to match OpenAI's exact wire format, which is what most local tooling integrates against, plus automatic speculative decoding for Qwen3.5 on Apple GPUs. None of this is a headline model, but it is the quiet part of local-first that decides everything: a capable open model is only useful once the runtime you already run can load it on the hardware you already have.

Cline, one of the most-used open-source coding agents, shipped four point releases in a week (4.1.3 through 4.1.6, August 2 to 6), and most of the fixes land squarely on BYOLLM setups. An empty model turn is now retried on every provider instead of only on Ollama, which had been a hard failure on OpenRouter and other OpenAI-compatible endpoints. A cold Ollama model load gets a 5-minute window to start responding rather than erroring out. Reasoning-effort settings now behave the same across providers, Ollama included. A run that exhausts its context window force-compacts and retries once instead of dying, and plan mode now hard-blocks file edits, redirects, and package installs rather than trusting the model not to make them. These are exactly the papercuts a local endpoint or a self-hosted model runs into first, and the tool is BYOLLM to the core, so you point it at whatever you already run.

DeepGrove released Maple-Preview, a 20B-parameter Mixture-of-Experts that runs at about 1B active and, unusually, is trained with ternary weights from the start rather than quantized down after the fact. The result is a 5.31GB checkpoint under an MIT license, a 131K-token context, and day-one quants for llama.cpp, Ollama, LM Studio, and Jan, reported at north of 200 tokens per second on an M4 Mac mini. DeepGrove claims leading reasoning for its weight class and 5 to 16 times the on-device speed of models like Gemma 4, Qwen3.5, and gpt-oss. Read the label: the authors call it a preview with minimal agentic post-training, so it is aimed at raw reasoning, not at driving an autonomous coding loop yet. But a native-ternary, MIT-licensed model that reasons well and fits in 5GB is a genuinely interesting take on making capable models small enough to actually own.

Liquid AI released LFM2.5-2.6B, a 2.69B-parameter model built for on-device work: a hybrid design (22 short-convolution blocks plus 8 grouped-query-attention blocks), a 128K-token context, and day-one GGUF, MLX, and ONNX builds so it drops straight into llama.cpp, Apple Silicon, or a browser runtime. Liquid reports about 220 tokens per second on an Apple M5 Max and roughly 30 on a phone, in under 2.5GB of memory, and a 77.83 on the ToolSandbox tool-use benchmark that edges out models several times its size. This is the runs-on-hardware-you-already-own end of local-first: small enough for a laptop or a handset, capable enough to drive tools. One thing to check before you build on it commercially: it ships under Liquid's own LFM1.0 license, not Apache or MIT, so read the terms rather than assuming.

Two weeks after its 975B Inkling, Thinking Machines Lab shipped Inkling-Small: a sparse Mixture-of-Experts with 276B total parameters but only about 12B active per token (each token routed to 6 of 256 experts plus 2 shared), Apache-2.0 licensed, and multimodal on input, taking text, images, and audio and returning text. Its self-reported SWE-bench Verified is 80.2 percent, with AIME 2026 at 95.5 and GPQA Diamond at 89.5. Treat vendor benchmarks as a starting point, but the shape is the story: 12B active is a fraction of what the full Inkling needs, and Apache-2.0 means you can pull the weights and use them commercially without asking anyone. One honest caveat, the same as its bigger sibling: 276B total is a server-class download, so "you can own it" and "you can run it on a laptop" are still two different claims.

Last week we noted that Moonshot had promised Kimi K3 for July 27 but that nothing was downloadable yet. Here is the correction: the weights went live on Moonshot's HuggingFace on July 27, within hours of that note, and a week on they are confirmed real. K3 is a roughly 2.8-trillion-parameter Mixture-of-Experts, about 104B active per token and multimodal, shipped ungated as 96 safetensors shards at 4-bit (around 1.6TB on disk) under a custom Kimi K3 license, and it is already among the most-downloaded open models of the week. Moonshot's own numbers put it near the top on coding and agentic tests, including a reported 88.3 on Terminal-Bench 2.1. Two honest caveats. A 2.8T checkpoint is a server-class download, so "you can own it" and "you can run it on your laptop" are not the same claim. And Alibaba's Qwen3.8 Max, teased months ago at a similar scale, still has no downloadable weights at all, while a rumored Qwen3.7 Flash drop has no primary source behind it. The rule holds: a model you can only rent is not one you own, and the ones worth building on are the ones with a file you can actually pull.

On July 31 DeepSeek released DeepSeek-V4-Flash, an MIT-licensed Mixture-of-Experts with a 1M-token context and a much smaller active-parameter count than its V4 Pro sibling, tuned for agentic and coding work (DeepSWE 54.4 and Cybergym 76.7 on the vendor's card). MIT is the part that matters: no usage strings, yours to pull, quantize, and serve. And the local stack caught up almost immediately. Inside the same week, DeepSeek-V4 support landed across the runtimes people actually run: llama.cpp added the architecture with multi-token prediction, ExLlamaV3 1.3.0 added preliminary support plus a CPU key/value cache tier, vLLM 0.26.0 shipped V4 performance kernels alongside native macOS arm64 wheels, and KoboldCpp 1.118.1 fixed multi-turn V4 chat. Community GGUF quants followed within days, with 3-bit builds reported to run in around 110GB of RAM. That is the local-first loop working as intended: a capable open model goes from weights-on-HuggingFace to runs-on-your-own-box in under a week, and with a BYOLLM setup nothing leaves your network.

A routine week of the cloud model menu rearranging itself under the people who depend on it. On July 31 GitHub Copilot deprecated Gemini 2.5 Pro and Gemini 3 Flash, pulling both out of its model picker, so anyone who had standardized on them has to move. A day earlier OpenAI recut prices on its cheaper GPT-5.6 tiers, with its Luna tier dropping about 80 percent, only three weeks after those tiers reached general availability at higher prices. Cheaper is welcome, but the same tier repriced twice in a month is a moving target you cannot plan a budget around, and a model that vanishes from a menu is one you never really controlled. This is the case for bringing your own model, in one week of headlines: the open weights you download do not get deprecated or repriced out from under you, and you switch providers on your terms, not the vendor's. Even Anthropic, a closed-model lab, published a post this week calling open-weights models without dangerous capabilities "a public good."

Two moves this week from the metered-cloud side of AI coding. On July 30 Tabnine, one of the older AI coding assistants, announced it has been acquired by Tricentis, an enterprise testing company that plans to fold Tabnine's context engine into its own quality platform. When a tool you depend on changes hands, its roadmap and pricing start serving the acquirer, not you. And on August 1 Replit cut prices across its cloud deployments and storage while leaning harder into tiered Agent models, Lite through Power, where the model you get depends on what you pay. Both are reminders of what a one-time, local-first, bring-your-own-model tool is built to avoid: no acquisition rug-pull, no metered cloud bill, no model gated behind a higher tier. You own the tool, and you pick the model.

Not every useful model comes from a lab you already follow. Kwaipilot, the coding team at Kuaishou, quietly shipped KAT-Coder-V2.5-Dev: a Mixture-of-Experts with 35B total parameters but only about 3B active per token, built on Qwen3.6, Apache-2.0 licensed, with a 262K-token context. Its self-reported SWE-bench Verified score is 69.4 percent, with Terminal-Bench 2.1 at 41.0 percent. Treat vendor benchmarks as a starting point, but the shape is what matters for local work: 3B active means it runs on modest hardware, and Apache-2.0 means you can use it commercially without asking anyone. A small, clean-licensed, coding-tuned model you can actually run is worth more to most developers than a trillion-parameter checkpoint they can only rent.

Open WebUI, the most widely self-hosted front-end for local models, shipped 0.11.0 on July 27 with sub-agents: background helper agents that run their own tool-driven conversations, alongside model-defined chat variables, shared-folder collaboration with read and write permissions, LDAP group sync, and a full UI redesign. The pattern from the last few weeks holds. The features that used to be the reason to reach for a cloud product, agents that plan and call tools on their own, keep landing in software you run on your own hardware. Point it at the models you already host and the whole loop stays on your network.

beta.33 fixes the check that tells you whether the agent's work actually worked. It had been treating "I could not observe this working" the same as "I watched this fail", so completed tasks came back marked failed and automated runs refused to apply them; measured against an independent grader across 89 real tasks, Bodega had solved 35 and wrongly reported failure on 29 of them. Now only an observed failure can fail a task, and the reverse case, a check that could not run outvoting one that really did fail, is fixed too. Claude Opus 5 joins the picker as a cloud model with a faster variant, 1M context, and automatic thinking behaviour. A set of safety rails built for small local models no longer second-guesses capable models on unattended work, you can cap how long a single run may take so it reports properly inside a CI job or test harness, OpenRouter gets a preferred-provider list that reuses context to cut cost and latency, and a dropped connection no longer runs your task twice. Still an open beta, free for everyone including commercial use, running on your machine with your own keys.

Two of the biggest open-weight promises in circulation still have not shipped a file you can pull. Moonshot's Kimi K3, the 2.8-trillion-parameter model announced on July 16 with weights promised for July 27, is not on Moonshot's HuggingFace organization as of today; a wave of aggregator posts claimed it dropped on July 26, but the primary source shows no K3 repository at all, so treat that as premature. Alibaba's previewed Qwen3.8 Max, teased at around 2.4 trillion parameters, is still cloud and API only, with no downloadable weights and no independent benchmarks. Both may still ship, and when the Kimi K3 weights land you will be able to pull and run them yourself. Until then the lesson is the one local-first keeps proving: an announced model you can only rent is not a model you own, and even a 2.8T checkpoint only helps if you have the hardware to load it. The models worth building on are the ones you can actually download and run today.

Poolside's Laguna S 2.1, the 118B open-weight coding MoE that runs at about 8B active, shipped its weights last week; this week the local runtimes caught up so you can actually run it. Ollama 0.32.4 (July 25) adds Laguna support on Apple GPUs through its MLX engine, along with smarter mixed quantization and a 4 to 9 percent decode speedup for Qwen3 MoE models on an M5 Max. ExLlamaV3 1.2.0, out the same day, adds Laguna support for NVIDIA cards plus experimental CPU offloading for expert layers, which matters because it lets a large Mixture-of-Experts model spill onto system RAM and run on a rig that could not otherwise hold it. A capable open coding model going from weights-on-HuggingFace to runs-on-your-own-GPU in four days is the local-first loop working as intended: point your BYOLLM setup at your own copy and nothing leaves your network.

Anthropic released Claude Opus 5 on July 24, positioned as near-Fable-5 intelligence at roughly half the cost, with a low, medium, high effort toggle and a 1M-token context. It holds the previous Opus price of $5 per million input and $25 per million output tokens, and is the new default on Claude Max. It is a closed, hosted model, so on its own it is another frontier step you rent rather than own. The BYOLLM angle is that there is no lock-in tax for using it: bring your own Anthropic key and it drops into a local-first tool the day it ships, right next to your on-device open models, with no new subscription and no switching cost. Bodega One Code v1.0.0-beta.33, out this week, already lists it in the picker. The closed frontier keeps moving; owning your tooling means you plug in whichever model wins each month without being tied to any one of them.

beta.32.1 lets the agent pick the right model for each step of a task on its own: a fast model to read and search, a code model for edits, a stronger model to plan and verify. It always shows which model handled what, and when a step fails the quality checks it retries once on a stronger model and tells you it did. It stays off unless you turn it on for an existing install, and is on by default for a fresh one. You can now also point Bodega at a model served on another machine on your network, though air-gap mode still refuses any non-local endpoint, so a remote model counts as an explicitly online mode. New in the picker: Qwen3-Coder-30B (2507), a vision-capable local Qwen3.6-35B, and the Qwen3.7-Plus cloud model. Still an open beta, free for everyone including commercial use, and it all runs on your machine with your own keys.

On July 21 Poolside released Laguna S 2.1, the bigger sibling to the 33B XS weights it opened earlier this month. It is a 118B-parameter Mixture-of-Experts (256 routed experts plus one shared, mixed global and sliding-window attention) that activates about 8B parameters per token, with a 1M-token context and full weights on HuggingFace under the permissive OpenMDW-1.1 license. It posts the strongest published open-weight numbers on several coding benchmarks: 78.5% on SWE-bench Multilingual, 70.2% on Terminal-Bench 2.1, and 59.4% on SWE-Bench Pro, matching or beating models many times its size. The BF16 checkpoint is a multi-GPU job at about 236GB, but the FP8, NVFP4, INT4, and GGUF quantizations shrink it enough to run on a single machine, and Poolside points to a single NVIDIA DGX Spark. The weights are yours to pull, quantize, and serve, so point a BYOLLM setup at your own copy and nothing leaves your network.

Ollama 0.32.3 (July 23) extends CUDA to Windows ARM64, so Snapdragon laptops get GPU acceleration, adds B200 support through CUDA 12, and lowers memory use on Linux CUDA and ROCm integrated GPUs. On the reliability side it fixes GLM tool calls being silently dropped at the end of generation, adds chat, thinking, and tool-calling support for the new Laguna 2.1 models with a Metal inference fix, and stops model downloads that stalled before sending data. Both the bundled MLX and llama.cpp engines were updated. If you run local models on ARM Windows or newer NVIDIA hardware, this is the release that turns the GPU on.

Two of the most-used local model apps shipped agentic and privacy features this week. Jan v0.8.4 (July 21) adds native web_search and web_fetch tools that run against the local models you already host, per-model chat-template kwargs for llama.cpp, token counting for remote and MLX providers, and it moves stored secrets out of localStorage into the operating system keyring. LM Studio 0.4.20 (July 22) adds LM Link, which serves your machine's models to its Bionic agent over your network, plus enterprise internal-network model endpoints and a developer log-level setting for the llama.cpp engine. The pattern from the last few weeks holds: the tools that run models on hardware you own keep growing into full agentic surfaces, so more of the loop stays off the cloud.

Another week of quiet speedups to the local stack. llama.cpp (builds b10094 through b10105, July 23 to 24) improved NVFP4 4-bit activation quantization on CUDA for Blackwell-class cards, made speculative decoding auto-infer the draft model type from the repo instead of a hand-set flag, optimized the Hexagon backend for Qualcomm NPUs on Snapdragon devices, and consolidated the mlock, mmap, and direct-I/O options into a single --load-mode argument. ExLlamaV3 v1.1.0 (July 18) cut CPU overhead in tensor-parallel mode and improved native-backend prefill for multi-GPU rigs. None is a headline on its own, but the tools that run models on hardware you already own keep getting faster, for free.

GitHub shipped a new Copilot usage-metrics dashboard on July 22 that buckets developers into adoption phases (Code-first, Agent-first, Multi-agent, Passive) and surfaces an "adoption multiplier" for organization admins. It is a small feature, but a telling one about where the seat-based tools are heading: more of the telemetry an employer sees about exactly how each developer works, tracked per seat. A one-time local tool with no account and no cloud round-trip has nothing to report in the first place.

Alibaba's top-tier Qwen model, Qwen3.7 Max, is turning up in more coding tools this week, including Bodega One Code's July 17 release, which added it to the picker alongside Kimi K3 and OpenAI's GPT-5.6 family. It is a capable 1M-token model available through Alibaba Cloud and resellers like OpenRouter at about $2.50 per million input and $7.50 per million output tokens. The catch for anyone here on privacy or cost grounds: the Max tier is closed-weight and API-only. There is no download, so it runs on Alibaba's servers at Alibaba's price, not on your machine. On July 19 Alibaba appears to have previewed the next step, Qwen3.8 Max, in a post from its Qwen account: a roughly 2.4-trillion-parameter model it calls its largest yet, and this time it says open weights are coming. That preview is not on the official model pages yet, and the numbers are Alibaba's own with no third-party coding benchmark behind them, so treat the whole thing as an unconfirmed announcement until it ships and someone independent runs it. The part worth watching is that open-weights promise. A frontier Qwen you can pull and run offline would matter a lot more than another flagship you can only rent.

On July 16 Moonshot unveiled Kimi K3, a 2.8-trillion-parameter Mixture-of-Experts model (896 experts, about 16 active per token) with a 1M-token context, and pitched it as an open-weight rival to Claude Opus 4.8 at a fraction of the price. At launch it posted the strongest published open-weight numbers on several agentic benchmarks: 93.5% on GPQA Diamond, 88.3% on Terminal-Bench 2.1, 91.2% on BrowseComp. The catch for anyone reaching for it on privacy or cost grounds is timing. The weights do not publish until July 27, so right now K3 runs only on Moonshot's API at a flat $3 per million input and $15 per million output. Until then it is another hosted frontier model on someone else's servers at someone else's price. When the weights drop you will be able to pull it and run it yourself. A 2.8T model wants serious hardware, but it will be yours.

LM Studio, one of the most-used apps for running models locally, released Bionic on July 16, a standalone agent for coding, research, and working with your files. What sets it apart from the hosted assistants is what it runs on: open models you host yourself or bring your own key for, like GLM 5.2 and Kimi K2.7 Code, rather than one vendor's cloud model. It does inline diffs, agentic search across a codebase, and local voice transcription through Voxtral, and it ships with a zero-data-retention commitment. The shift here is that the local-inference tools are no longer just model runners; they are growing full agentic coding surfaces, so more of the loop that used to require a cloud subscription can run on hardware you already own. It is also a sign the local-first coding space is filling in fast. Host the models yourself and the work stays on your machine.

Another week, another round of quiet speedups to the local stack. llama.cpp (build b10042, July 16) turned on CUDA graphs for Volta and Turing GPUs, so budget and second-hand NVIDIA cards decode noticeably faster. On the server side, vLLM 0.25.0 (July 11) made its Model Runner V2 the default for dense models and brought the Transformers backend to performance parity, and SGLang 0.5.15 (July 14) added smarter Mixture-of-Experts load balancing for DeepSeek-class models plus NVFP4 support for Qwen3.6. None is a headline on its own, but the pattern holds: the tools that run models on hardware you already own keep getting faster, for free.

Thinking Machines Lab released Inkling on July 15, its first open-weight model and a serious one for agentic coding. It is a 975B-parameter Mixture-of-Experts (a 66-layer decoder routing each token to 6 of 256 experts plus 2 shared) that activates about 41B parameters per token, with a 1M-token context and full weights on HuggingFace under Apache 2.0. It reports 77.6% on SWE-bench Verified, frontier-class among open weights. At 975B total this is a multi-GPU or high-VRAM job to run locally, not a laptop model, but the license is clean for commercial use and the weights are yours to download, quantize, and serve. Point a BYOLLM setup at your own copy and nothing leaves your network.

A community fine-tune trending on HuggingFace this week takes the reasoning tax off local coding. ThinkingCap-Qwen3.6-27B, from BottleCap AI, retrains Qwen3.6-27B to stop reasoning once it is confident, cutting thinking tokens by about half on average and up to 90% in the best cases while accuracy holds flat or improves. On LiveCodeBench it rose from 80.7% to 84.3% while spending 41% fewer thinking tokens; on GSM8K, 93.3% to 96.5% with 74% fewer. Fewer tokens per step means lower latency and lower cost on every local agent loop, which matters more when you are paying in your own GPU seconds rather than a provider's meter. Apache 2.0, with GGUF builds for llama.cpp, so you run it yourself.

On July 15 Anaconda acquired Kilo Code, the open-source AI coding extension with more than 3 million developers. Per the announcement Kilo stays open-source and self-hostable, now paired with Anaconda's package tooling and an MCP router that surfaces per-model token cost. It is a softer landing than the Continue.dev acquisition that shut that product down a month earlier, but the lesson is the one we keep flagging: when the tool you depend on is someone's company, its future is a boardroom decision, not yours. The parts that survive an acquisition intact are the ones you can run yourself, open weights, a local runtime, an editor you own. Kilo staying open-source is exactly the thing that makes this one land softly.

PrismML released Bonsai 27B on July 14, a compression of Qwen3.6-27B to extreme low-bit weights that still holds most of the original quality. The ternary build (1.71 bits per weight) is 5.9 GB and the 1-bit build (1.125 bits) is just 3.9 GB, small enough to run a 27B-class coding model on a laptop or even a phone, with the full 262K-token context intact. On a combined coding suite the ternary variant scores 86.0 and the 1-bit 81.9 against 88.7 for the full-precision baseline, so you trade a few points for an order-of-magnitude smaller footprint. It ships Apache 2.0 with native builds for Apple devices via MLX and NVIDIA GPUs via CUDA. This is the local-first thesis in one release: real coding capability on hardware you already carry, with nothing leaving the device.

Anthropic ended included Claude Fable 5 access on its paid Claude subscriptions. After a deadline extended to July 12, Fable 5 shifts to prepaid usage credits at $10 per million input tokens and $50 per million output tokens - the highest rate Anthropic has published for a generally available model. Included subscription access is meant to return when capacity allows, but for now, without credits enabled, access stops with no grace period. If you drive Fable 5 through a bring-your-own key from a tool like this one, that is the per-token rate you pay now. When your coding runs on a hosted frontier model, the price and the terms stay theirs to change.

Ollama 0.32.0 (July 11) changed what happens when you type bare "ollama": instead of dropping into a model chat, it now launches an interactive agent that can chat, write code, search the web, and delegate tasks. The most-installed local runtime is moving from serving models to running agentic loops itself, onto the same ground as Cline, Aider, and OpenCode. The release also began deprecating older model tags such as CodeLlama and Qwen2.5-coder in favor of current ones. For a local-first setup it means one more capable agent that runs entirely on your machine and your weights, with no account and no per-token meter.

OpenAI made GPT-5.6 Sol generally available on July 9, the top tier of a new family (Luna, Terra, and Sol from least to most capable). OpenAI calls it its best coding model yet: it tops the Artificial Analysis Coding Agent Index at 80, about 2.8 points above Claude Fable 5, while using less than half the output tokens and costing roughly a third less per task. Sol is OpenAI-hosted only, running in ChatGPT, Codex, and the API at $5 per million input and $30 per million output; Terra and Luna are cheaper. It is a genuinely strong coding model and, like every hosted flagship, one you rent rather than own. The price, the availability, and the retention terms stay OpenAI's to change. The BYOLLM point is not that local weights beat it today, it is that a model on your own hardware cannot be repriced or switched off underneath you.

Bodega One Code v1.0.0-beta.31.10.3 makes its model catalog self-updating. Instead of waiting for an app update to see new models, Bodega can refresh the catalog (model list, hardware-fit info, and downloadable GGUF entries) from a single hardcoded Bodega URL. It fetches one JSON file and nothing else - no telemetry, nothing about your machine is sent - is on by default with a Settings toggle, and is disabled entirely in air-gap mode where Bodega never reaches out. The July catalog refresh adds Grok 4.5 (SpaceXAI, an Opus-class cloud model at $2 / $6 per 1M tokens) and GLM-5.2 (Zhipu, a 1M-token MIT open-weight coding MoE) to the cloud list with pricing, corrects the Kimi context length, and adds two local vision models with their image projector files so they run out of the box: Qwen3-VL 32B and MiniCPM-V 4.5. The Discover screens also group models by family now, show capability tags (tools, vision, thinking, FIM, draft, MoE), and tell you whether a model fits your GPU, and which quantization fits best, before you download. Bring your own key; it all runs on your machine.

Bodega One Code v1.0.0-beta.31.10.2 teaches its downloader to handle multi-file (sharded) GGUFs, which unlocks Qwen3-Coder-Next 80B - the 80B / 3B-active MoE that tops SWE-bench Verified around 71% among locally-runnable coding models. Bodega now fetches every shard, resumes each one independently if the download drops, loads them as a single model, and deletes them together on removal (Q4_K_M is about 45 GB across four files, so a 48GB+ card). SmolLM3-3B and Mistral Small 3.2 24B were added and verified against HuggingFace, a false "size mismatch" that had been rejecting complete GGUF downloads (including Qwen3.6 27B) was fixed, and a local model default context is now sized against the card total VRAM instead of whatever is momentarily free. Bring your own hardware; nothing leaves your machine.

Zed 1.10.0 (July 8) added llama.cpp as a first-class language-model provider. You can point the Zed agent at a local GGUF straight away, with no separate inference server or OpenAI-compatible shim in between. The same release folded LLM providers, external agents, and MCP servers into the main settings editor and extended reasoning-level controls for OpenAI-compatible and OpenCode backends. For anyone who wants the editor agent to run fully on their own machine and weights, this removes a step.

Three local inference runtimes updated in the same week. Ollama 0.31.2 (July 6) turned on flash attention for older NVIDIA cards (compute capability 6.x) and let integrated GPUs offload vision models, so more modest hardware runs faster. LM Studio 0.4.19 (July 7) made its Engine Protocol the stable default and fixed reasoning-model output being replayed as ordinary content through the responses endpoint. MLX 0.32.0 (July 7) added nvfp4 and mixed-bit quantization plus CUDA and Metal kernel work, for better local inference on Apple Silicon and beyond. None is headline-grabbing alone, but the local stack keeps getting faster on the hardware people already own.

Tencent released Hy3 on July 6, an open-weight Mixture-of-Experts model under a clean Apache 2.0 license. It is 295B total parameters but activates only about 21B per token (192 experts, top-8) with a 256K context window, so it runs at roughly 21B-active compute while rivaling flagship open models several times its active size. Full weights are on HuggingFace day one (about 598 GB, or roughly 300 GB for the FP8 build), and community quantizations for llama.cpp, Ollama, and LM Studio followed almost immediately, so a high-VRAM or multi-GPU local rig can run it. It is tuned for agentic and coding work and reports around 78% on SWE-bench Verified. Apache 2.0 means no strings for commercial local use. Point a BYOLLM setup at it and run it yourself.

Poolside, the AI-coding company, released Laguna XS 2.1 on July 2, its first open-weight model and one built specifically for agentic coding. It is a 33B Mixture-of-Experts that activates about 3B parameters per token (40 layers, 256 experts) with a 262K context window, so it runs on a single GPU, and the weights ship on HuggingFace in BF16, FP8, NVFP4, and INT4. The license is the notable part: OpenMDW-1.1, a fully permissive license designed for model weights, which keeps it clean for commercial local use. Poolside reports a 5.4-point jump to 63.1% on SWE-bench Multilingual over the prior Laguna XS.2. Another capable, permissively licensed coding model you can point a BYOLLM setup at and run yourself.

InternScience (Shanghai AI Lab) open-sourced Agents-A1 in late June, a 35B Mixture-of-Experts model under Apache 2.0 built for agentic and tool-use work. In the first week of July, quantized GGUF builds landed for llama.cpp, Ollama, and LM Studio (53 variants in all), so it runs on a single consumer GPU, and it climbed into Hugging Face top trending. The model card reports strong agentic and instruction scores (79.81 on tau2-Bench, 44.33 on SciCode, 94.82 on IFEval) and argues its edge comes from training for long task horizons rather than raw parameter count. For BYOLLM it is another permissively licensed model that runs on your own hardware and holds its own on agentic coding.

On July 1, GitHub made Kimi K2.7 Code the first open-weight model selectable in the Copilot model picker, rolling out to Pro, Pro+, and Max first (off by default for Business and Enterprise). The catch for anyone reaching for open weights on privacy or cost grounds: it runs on GitHub Azure servers, not your machine, and bills under usage-based pricing at provider list rates. A day later Copilot added AI-credit pools so enterprises can cap how much of the monthly credit allotment each cost center spends. The BYOLLM contrast is the whole point: the same open weights run locally through Ollama or LM Studio cost nothing per token and never leave your machine.

Anthropic brought Claude Fable 5 back to general availability on July 1, three weeks after suspending it worldwide on June 12 under a US export-control directive. The US Department of Commerce lifted the order on June 30 once Anthropic agreed to a set of security commitments and shipped an improved safety classifier that blocks the reported jailbreak and routes flagged requests to the older Opus 4.8 model. Fable 5 is back across the Claude platform and Amazon Bedrock, with Google Vertex AI and Microsoft Foundry to follow. Mythos 5, suspended alongside it, is only partially restored for a limited set of US organizations, not worldwide.

Anthropic released Claude Sonnet 5, a new flagship in the Sonnet line with a 1M-token context window by default and up to 128k output tokens. Anthropic positions its performance near Opus 4.8 at a much lower price, with introductory API pricing of $2 per million input tokens and $10 per million output through August 31, then $3 and $15. It is now the default model on the Free and Pro plans. For BYOLLM setups the takeaway is a near-frontier Claude model at a fraction of Opus cost that you can point your own key at.

beta.31.7 adds Claude Sonnet 5 and brings Claude Fable 5 back to Bodega One Code, both available through the Anthropic provider, OpenRouter, and Cloud Boost. Repeated turns on local models are faster now that llama.cpp and Ollama reuse the prompt cache and keep the model warm between turns. The release also fixes a Cloud Boost stall on Claude models driven off a local base, keeps the full step budget for capable models on long tasks, and clears a batch of preview, model-switching, and licensing bugs.

DeepSeek said its official V4 release lands in mid-July with a 1M-token context window across the lineup and stronger agentic, math, and code performance. It also introduces time-based API pricing, a first for DeepSeek: requests during peak hours (9am to 12pm and 2pm to 6pm) are billed at twice the off-peak rate. A peak-hour surcharge on a metered API is a new twist on renting your inference, and a clean argument for owning it. Running V4 weights locally once they ship sidesteps the meter entirely: no per-token rate, no midday multiplier, no clock-watching.

beta.31.6 adds Bodega Mixture, an optional Mixture-of-Agents engine: several reference models run in parallel on your turn and one aggregator model synthesizes the single reply. The reference models run with no tools and the conversation text only; the aggregator owns the tools and writes the answer. Local reference models are near-free, so an all-local mixture costs about one cloud call, and an optional QEL quality gate can verify the output. It is off by default and honors air-gap (cloud references drop when you are offline). The release also lets the in-app Preview serve plain static sites and games with no dev server, on loopback only, plus fixes for long Cloud Boost builds and a local-model context-window clamp.

llama.cpp build b9840 merges DeepSeek V4 graph optimization and GGUF conversion tooling, so the MIT-licensed V4-Pro and V4-Flash weights can now be converted and run locally as GGUF. Build b9833 added a MiniCPM5 parser with XML tool-call support, and b9837 added a reasoning-preserve option for chat templates. This is the unglamorous substrate that decides whether a new open model is something you can own and run, or just read about.

LM Studio 0.4.17 adds prompt-template overrides when loading GGUF models, load-time speculative decoding, support for AMD Strix Halo and Radeon AI PRO GPUs, and Mermaid diagram rendering in chat. 0.4.18 follows with interface-scaling and reasoning-model title fixes. The Strix Halo support matters if you are eyeing one of the new unified-memory mini PCs as a local-inference box: more hardware that runs models on your desk, not someone else's.

Ollama 0.30.11 unifies and tunes speculative decoding across its MLX runner, corrects inverted integrated and discrete GPU classification on Windows hybrid-graphics laptops, and stops double-counting memory-mapped weights in ollama ps. It also auto-installs Claude Code and opencode when they are missing. Routine plumbing on the runtime most BYOLLM setups lean on.

Liquid AI released LFM2.5-230M, a 230-million-parameter open-weight model aimed at on-device inference. It ships in the formats that matter for running locally, GGUF for llama.cpp and MLX for Apple Silicon plus ONNX and vLLM, and is small and fast enough to run on a phone or a Raspberry Pi rather than a workstation. It is not a frontier coder; the point is the other end of the spectrum, where a capable model runs entirely on hardware you already own, with no endpoint and no per-token bill.

Sakana AI launched Fugu, a multi-agent system "delivered as one model": a trained orchestrator that routes each task across a proprietary pool of frontier models behind one OpenAI-compatible endpoint. Sakana reports 73.7 on SWE-bench Pro for Fugu Ultra and says its Fugu models surpass publicly accessible frontier models. It is closed and API-only, with no open weights, and is not offered in the EU or EEA. Worth setting next to our own Mixture feature, which runs the same Mixture-of-Agents idea the opposite way: locally, across your own models, with the coordination on your machine instead of your code going to a vendor cloud.

Continue, the open-source BYOLLM coding extension for VS Code and JetBrains, has been acquired by Cursor. Its repository is now read-only after a final v2.0.0 release on June 19, and continue.dev says the mission continues through Cursor. The Apache-2.0 code stays available for anyone to fork, but there are no further updates from the original team. It is the same lesson as the Gemini CLI shutdown and the Fable 5 suspension, one more time: a tool you depend on can be bought, re-gated, or wound down on someone else's schedule.

The Gemini-CLI-into-Antigravity migration Google announced at I/O now has firm shutdown dates, and they only spare enterprise. Google is discontinuing Gemini CLI and Gemini Code Assist for consumer accounts: the shutdown begins June 18, 2026 and completes July 17. Consumer users get moved to the new Antigravity CLI, which Google itself says does not have 1:1 feature parity at launch. Enterprise keeps access. It is the second time in a week that a hosted coding tool got pulled or re-gated on the vendor's timeline (Anthropic disabled Claude Fable 5 worldwide on June 12). The throughline for anyone building on a hosted endpoint: it can be deprecated out from under you. A model and an editor you run locally keep working until you decide otherwise.

The beta.29.2 through beta.29.5 point releases are a reliability and privacy pass on the beta.29 routing line, and one of them is a catalog refresh worth its own entry. Five current GGUF models join the llama.cpp catalog so you can run them locally: Qwen3-Coder-30B, Gemma 4 12B, DeepSeek-R1-0528-Qwen3-8B, Qwen3.5-9B, and Qwen3-VL-30B. The managed llama-server binary moved to build b9670. The rest of the line is bug fixes, openly, because this is a beta: a crashed llama-server recovers instead of wedging the app, downloaded GGUFs list from disk before the server is running, there is no more silent fall-back to Ollama once you pick a local model, and a stale model id from an old Ollama setup no longer jams startup. beta.29.5 also makes the Code-mode editor fully air-gapped by shipping its Monaco bundle inside the app instead of fetching it from a CDN, so opening the editor under air-gap makes zero outbound requests, and it hardens beta sign-up to queue your contact on-device before contacting Loops so a network error no longer drops it.

Z.ai released GLM-5.2 weights on Hugging Face under an MIT license, the successor to last month's GLM-5.1. It scores 62.1 on SWE-bench Pro against GPT-5.5's 58.6 (GLM-5.1 was 58.4), the strongest open-weight coding model on that benchmark, with a 1M-token context window and 26 community quantizations, with support for llama.cpp, Ollama, LM Studio, and Jan. Same honest caveat as 5.1: at 753B parameters this is a workstation or multi-GPU model (roughly 8x H100), not a laptop or single-24GB-card model. The point is not frontier coding on your MacBook. It is that a frontier-tier coder is now something you can own and run on your own hardware, air-gapped, under a permissive license, instead of renting it from an endpoint that can change or vanish.

The local runtime most BYOLLM setups lean on shipped a run of releases this week (0.30.8 on June 12 through 0.30.10 on June 17). The change that matters for agentic coding: prompt caching is now decoupled from context shifting for better KV-cache reuse, the same within-session token-reuse win that keeps a long local agent run from re-reading its whole history every turn. Also in: a fix for ollama launch picking the wrong provider in some cases, and steadier MLX inference on Apple Silicon. Unglamorous plumbing, but it is the substrate that makes bring-your-own-local-model fast enough to actually live in.

beta.29 is the routing release. You write an ordered, first-match-wins rule table (Settings, AI Behavior, Routing) that decides which model handles what, matching on mode (chat or code), ask type, agent step (read, write, plan, verify), file path, message size, or how much you have spent today, then routing to a specific model or keeping it local. OR and NOT conditions, custom classifier patterns, per-project rules shipped in .bodega/config.json, an Auto-pill preview, and a dry-run tester are all in, and hard limits still clamp every decision so a rule can never bypass air-gap, VRAM limits, or spend caps. Verified Private Automation points Bodega at a plain task or a GitHub issue and does the whole job on your machine: it works in an isolated git worktree, writes the code, runs full QEL verification (boots your server and probes a real route, runs your tests, checks the contract), and opens a pull request with the verification trace in the description. A passing run gives a ready PR; a failing one opens a draft. Cursor and Copilot do this by shipping your code to their cloud; Bodega runs it on your hardware and nothing leaves your machine but the branch and the PR. The agentic loop got smarter too: Goals (/goal) that survive across messages and send a second model in to attack the result before a goal can complete, after-edit type diagnostics from a bundled language server for TypeScript, JavaScript, and Python, a dispatch_scout read-only sub-agent that explores the repo without filling your context, a get_diagnostics tool, and test-driven repair when a build fails verification twice. The editor language server also graduated from experimental and now ships in the app for everyone, cost tracking covers your own self-hosted models, sessions export as a self-contained web page or a self-hostable viewer, and a catalog refresh adds fourteen models and two cloud providers (MiniMax and Z.ai). Claude Fable 5 was removed from the catalog after Anthropic withdrew it from general availability.

Three days after Fable 5 went generally available, Anthropic took it offline. On June 12 the company said it received a US government export-control directive that afternoon, citing national security authorities, and had to abruptly disable both Fable 5 and Mythos 5 for all customers to comply. The order targeted access by any foreign national, inside or outside the United States, which left a global shutoff as the only practical way to enforce it. Anthropic said it disagreed with the directive, arguing that a narrow potential jailbreak should not be cause to recall a commercial model used by hundreds of millions of people, but complied anyway. Access to its other models, including Opus 4.8, was unaffected. The policy debate aside, the mechanism is what matters here: a hosted model can be switched off for everyone, on someone else's timeline, with no notice. Weights you already run on your own hardware cannot. That is the case for BYOLLM and local-first in a single news cycle.

Moonshot AI releases Kimi K2.7-Code, its fifth Kimi model in under a year, open-weighted on Hugging Face under a Modified MIT license and reachable through the Kimi API and the Kimi Code CLI. It keeps the same 1T-parameter mixture-of-experts architecture as K2.6 but retrains the reward model and data pipeline around long-horizon, real-world software tasks. The headline is efficiency rather than a leaderboard score: Moonshot reports roughly 30% fewer reasoning tokens than K2.6 and a 21.8% gain on its own Kimi Code Bench v2. Every launch benchmark is first-party (Kimi Code Bench v2, Program Bench, MCP Atlas and friends); no third-party SWE-bench Verified, SWE-bench Pro, or Terminal-Bench numbers were published at launch, so the public-suite picture is still open. The catch for local-first work: the weights run to roughly 600GB at full precision (around 240GB heavily quantized), with no official GGUF or Ollama build yet, so this is server-class, not laptop-class. Open license does not mean it loads on your machine. BYOK reaches it today; running it yourself needs real hardware.

Kiro, Amazon's agent-first IDE, shipped a cluster of updates in the second week of June. June 11 brought GitLab support to the web app (connect with a personal access token, then clone, edit, and open merge requests from chat) and moved the Specs workflow into the browser, so the plan-before-code pass of requirements, design, and tasks now runs without the desktop app. June 10 added a Pro Max tier at $100 per month, slotted above Pro+. June 12 followed with CLI 2.7.0: a /goal command for iterative goal-driven loops and Queue Steering to redirect an agent mid-turn. Kiro is cloud-hosted and metered rather than local, but its Specs-first flow maps onto the same plan-and-verify discipline a local agent can run on your own machine.

One day after beta.27, beta.28 ships the polish pass: a 25-agent research sweep over everything already shipped, turned into five waves of fixes and the small features users kept asking for. Run Tasks puts one-click dev-server launch in the terminal tab bar: it auto-detects dev/start/serve/watch scripts from package.json, takes custom named tasks via a run_configs key in .bodega/config.json, runs each task in its own terminal tab, and offers npm install when node_modules is missing. The Runs Inbox is one topbar surface for everything waiting on you (fleet sessions needing approval or ready to apply, loop runs parked for review or below their QEL bar), hidden when empty. Switching from another editor got cheaper: opening a project with Cursor, Copilot, Cline, Windsurf, or Continue rules offers a one-click append-only import into .bodega-rules. Ask mode can now auto-approve read-only tools (off by default; file writes, web calls, and shell always ask, and shell can never auto-approve). Cloud models show a pre-send cost estimate in the status bar; local models show nothing because there is nothing to pay. The source-control Review button re-reviews only files that changed since the last pass, loops show a live next-run countdown with a per-loop Run-now button, and every llama.cpp model swap shows staged progress. The fixes worth naming: cost tracking showed $0.00 for Claude Opus 4.8 and Fable 5 because the pricing table was missing both flagships; the air-gap warning about configured cloud providers never fired; long-lived ACP connections leaked memory; and quitting mid-run could kill headless loops mid-write. Performance and consistency round it out: Map staleness checks hash 16 files at a time, llama.cpp loading no longer blocks boot, one focus ring across 35 controls, and status colors finally read correctly on light themes.

Cursor rebuilt Bugbot on its Composer 2.5 model and says the review agent is now over 3x faster, 22% cheaper, and finds 10% more bugs per run, with 90% of reviews finishing in under three minutes. A new /review command runs Bugbot and Security Review inline before a push, with /review-bugbot and /review-security to call each one directly, and it syncs with GitHub and GitLab to skip re-reviewing identical diffs. Bugbot is the closest competitor surface to Bodega One Code's QEL, with one structural difference: Bugbot reviews your diff in Cursor's cloud, while QEL runs its checks, including booting your server and hitting a real route, on your machine.

Zed's v1.6.3 release adds a Fast mode toggle in the agent panel that trades reasoning overhead for speed on Anthropic and OpenAI models, and overhauls terminal sandboxing so an agent requests write access to specific paths per command instead of demanding the whole filesystem. Skills can now be shared by link and loaded from symlinked global directories, and the Git panel gains split diffs in history and added/removed line counts. Zed matters here because it runs Bodega One Code's agent over ACP as of beta.25: tighter per-command permissions on the host side fit the same sandboxed, least-privilege model Bodega enforces in-app.

beta.27 makes the agent run on a schedule and makes "verified" mean the code actually runs. Bodega Loops (Automations) define a named agent task that fires on a cron schedule, a fixed interval, or Bodega Map staleness (when N of your codebase-map module summaries go stale). Every run executes headless in an isolated git worktree and lands as a reviewable diff with its QEL score and full trace. Park-for-review is the default; auto-apply only ever happens when the QEL score clears the bar you set; dry-run never applies anything. Each loop picks its own model (a custom agent, or a specific provider plus model), so a fast local model can take the small jobs while a cloud model handles the heavy ones. Scheduled runs respect spend caps, inherit air-gap, cap their own iterations, and refuse to run without worktree isolation. The verification layer got a ground-up overhaul. QEL execution proofs now boot the generated server in the same sandbox, force an ephemeral port, and send one request to the first route the task asked for, so a response under 500 is the strongest pass evidence QEL has while a boot crash or a 5xx is a real failure (loopback only, secrets-free env, 12-second cap, process tree always killed). A 43-scenario calibration harness measures the verifier against labeled known-good and known-bad work, with CI floors that fail the build if broken code passes; the first sweep caught and fixed two real scoring holes. The semantic judge now runs air-gapped on a localhost model and can park a marginal pass for human review. Verification also runs once instead of up to four times, with proof gates in parallel, for roughly half the wall-clock. Beyond QEL: the terminal gains a sticky running-command header, a real Output panel, and a real Debug Console; Bodega registers as a file handler so double-clicking a file opens it; hunk-level staging and a conventional-commit type picker land in source control; and the local API server now refuses to start under air-gap, closing a no-network hole.

Anthropic ships Claude Fable 5 into general availability across the Claude API, Amazon Bedrock, Vertex AI, and Microsoft Foundry. It runs a 1M-token context window by default with up to 128k output tokens per request, priced at $10 per million input tokens and $50 per million output. One detail matters for privacy-sensitive work: Fable 5 carries 30-day data retention and is not offered under zero data retention, so prompts and completions sit with the provider for up to a month. Subscribers on Pro, Max, Team, and Enterprise got it free through June 22; after that, continued use bills usage credits at API rates. A capable model, and a clean reminder that BYOK plus local models is how you keep the sensitive work off someone else's servers.

Cohere Labs releases North Mini Code 1.0 on Hugging Face under Apache 2.0. It is a 30B mixture-of-experts model with about 3B active parameters, small enough to run on a single H100, and Cohere reports roughly 80% pass@10 on SWE-bench Verified, ahead of open models several times its size. For local-first coding it is the interesting kind of release: an openly licensed, genuinely small coding model you can host yourself instead of renting one by the token.

beta.26 closes the bidirectional ACP loop. Beta.25 made Bodega an ACP host (Cursor, Claude Code, Gemini CLI, Codex run inside the fleet); beta.26 makes it an ACP server too. `bodega --acp` starts Bodega headless so the agent can be driven over stdio JSON-RPC by an external editor; register it in Zed's `agent_servers` and Bodega's agent runs inside Zed with the same permission gates and `acp.allowed_projects_dir` confinement as in-app. Custom agents land as a Settings profile (system prompt + pinned model + tool allowlist + read-only flag + iteration cap), with the allowlist enforced at the execution gate as an intersection with the panel's tools so a profile can only narrow what runs, never widen it. The Bodega Map exposes itself to the agent as a callable `query_map` tool the agent invokes mid-task for grounded answers with sources. Plus: local code review on the working diff, per-hunk diff accept/reject from the Agent panel, Claude Fast mode toggle, managed llama.cpp embedding server, hard cloud-spend caps, an Air-Gap Active top-bar indicator, terminal Ctrl/Cmd-click code links, an in-app docs hub, and a Settings restructure into seven labeled groups. Hardening: Windows backend tree-kill + parent-death watchdog fixes the orphan-port-3000 bug; security pass returned zero critical and zero high; air-gap extended to Git AI features so a commit-message or review request cannot leak a diff to the cloud.

Bodega One Code becomes an orchestration platform. Three landmark features ship together. Fleet Parallel fans one task out to multiple models in isolated git worktrees, then QEL scores the resulting diffs so you pick the winning attempt; watching three models try the same refactor and diffing the output is the feature most users hit first. ACP (Agent Client Protocol) support lands: Cursor, Claude Code, Gemini CLI, and Codex now run inside Bodega's fleet, routed through Bodega's sandboxed filesystem and shell so air-gap mode still holds. The Codebase Wiki wires the dependency-graph view into the embedding index Bodega was already building, producing a queryable knowledge layer that can explain a file, document the repo, or answer questions grounded in real code. Pre-ship: a 6-agent review fleet (security, architecture, perf, tests, docs) ran in parallel; 6,203 backend + 1,682 frontend tests green at tag.

Cursor ships four updates in the first week of June. June 5: Design Mode in the Cursor browser lets you click, draw, or describe UI changes by voice while an agent is mid-run. June 4: SDK update lands custom stores, user-defined tools wired through a built-in custom-user-tools MCP server, optional auto-review on local tool calls, and nested subagents at any depth. June 3: Enterprise Organizations goes GA, letting customers manage multiple Cursor teams under one umbrella with per-team security, governance, and budget controls. June 1: Teams pricing splits usage pools, adds a new Premium seat for heavy agent users, and improves spend forecasting.

VS Code 1.122 lands org-governed Copilot plugins in public preview: admins push agents, skills, and MCP servers to a whole team through settings.json. Alongside it, GitHub adds a bring-your-own-key path that drops the GitHub sign-in requirement, aimed at air-gapped and restricted environments. It is a notable step onto local-first turf, though the agent still runs through Copilot and the cloud account model rather than end to end on your own machine.

NVIDIA ships Nemotron 3 Ultra, a 550B-parameter mixture-of-experts model (about 55B active) with a hybrid Mamba-Transformer architecture and a 1M-token context, released under the permissive OpenMDW-1.1 license with weights, training data, and recipes all published. It is a datacenter-scale model rather than a laptop one, but it pushes the open-weight frontier and gives self-hosting teams a fully inspectable flagship. Independent analysis still places it behind the leading Chinese open models like Kimi K2.6.

Google adds a 12B model to the Gemma 4 family under Apache 2.0, sitting between the edge-friendly E4B and the 26B mixture-of-experts model. It is the first mid-sized Gemma with native audio input, handling text, image, and audio, and Google says it runs locally in about 16GB of VRAM or unified memory, so it fits a typical laptop. With the Gemma 4 line past 150 million downloads, it is one of the more practical open multimodal options for local work.

Cognition ships Devin Desktop as a standard over-the-air update to Windsurf on June 2. The Agent Command Center becomes the default surface: every local and cloud agent gets managed from a single Kanban view. New Spaces feature groups sessions, PRs, files, and context so agents can share state. The legacy Cascade agent gets replaced by Devin Local, rewritten from scratch in Rust, up to 30% more token-efficient, with native sub-agent support. Cascade stays available through July 1 for incremental migration. Devin Desktop ships with Agent Client Protocol (ACP) support at launch, accepting Codex, Claude Agent, OpenCode, and any other ACP-compatible agent including in-house ones. All Windsurf settings, plans, pricing, extensions, keybindings, and MCP connections port over automatically.

Anthropic publishes full documentation for Dynamic Workflows, the Opus 4.8 flagship orchestration mode that landed at the May 28 model release. In a single Claude Code session, Opus 4.8 writes an execution script and spawns up to 1,000 parallel subagents to complete tasks. Currently in research preview on Max, Team, and Enterprise plans. Pairs with Claude Managed Agents capabilities including dreaming, multiagent orchestration, outcomes, and webhooks. Higher Claude Code and Opus API rate limits ship alongside the docs.

One day after the AI Credits billing flip, The Register reports developer accounts of bills landing harder than the rate table suggested. The highest-resonance reports are single-call burns rather than monthly depletion: one Pro+ user posted a 1,180-credit ($11.80) Claude 4.8 website-editing session, 16% of the monthly pool, on a single interaction; another reported $6 in credits charged on one project change request. The GitHub community discussion #192948 sits at 958 down vs 24 up reactions with the thread locked on June 1. The most-cited migration alternatives in the thread: Cursor Pro, Devin Desktop, Roo Code (archived May 15), Cline plus OpenRouter, and direct Anthropic API.

At Build 2026, Microsoft announced MAI-Code-1-Flash, its first in-house AI coding model, and began rolling it into GitHub Copilot for individual users in VS Code through the model picker. Microsoft says it was trained on clean, traceable, enterprise-grade data without distillation from third-party models, and reports a 16-point lead on SWE-Bench Pro over Claude Haiku 4.5 (51.2% versus 35.2%) while using up to 60% fewer tokens on SWE-Bench Verified. The model is proprietary and Copilot-only; there is no open weight and no way to run it on your own machine. It is a clear direction signal: Copilot is consolidating around a vertically integrated stack built on Microsoft's own model, the opposite of a bring-your-own-model approach.

MiniMax announces M3 as the first open-weight model to combine frontier-level coding, a 1-million-token context window, and native multimodal capabilities in a single architecture. Scores 59.0% on SWE-Bench Pro (the harder agentic-coding benchmark), surpassing OpenAI GPT-5.5 and Google Gemini 3.1 Pro on the same test (trails Claude Opus 4.8 at 69.2%). Also posts 66.0% Terminal Bench 2.1, 34.8% SWE-fficiency, 28.8% KernelBench Hard, and 74.2% MCP Atlas. The novel MSA (Multi-Stack Attention) architecture cuts per-token compute at 1M context to one-twentieth of the prior generation, with 9x faster prefill and 15x faster decoding. API access live at launch; open weights and the technical report ship to Hugging Face and GitHub within ~10 days.

Bodega One Code extends its two-model vision orchestration from Ollama to llama.cpp, the managed local engine. A local text-only coding model (DeepSeek, Qwen-Coder, Llama-3.3) can now answer questions about an attached image: Bodega hot-swaps in a bound local VLM, answers against the image, and swaps the coding model back in a single round trip while you stay in context. Cloud models handle vision natively, so they are untouched. Also ships a guided-tour overhaul, an in-app changelog, and llama.cpp onboarding hardening. Beta extended to November 1.

The June 1 cutover is live. Paid plans now bundle base credits matched 1:1 with subscription price plus a flex allotment: Pro ($10/mo) ships $10 + $5 = $15 in premium-request credits; Pro+ ($39/mo) ships $39 + $31 = $70; the new Max plan launches at $100/mo with $100 + $100 = $200. A heavier model burns more credits per call. GitHub paused new Pro, Pro+, and Max sign-ups at the cutover "to ensure a high-quality experience"; existing customers can still upgrade.

Anthropic releases Claude Opus 4.8, 41 days after Opus 4.7 (a much faster upgrade cycle than usual). Agentic coding score climbs from 64.3% (4.7) to 69.2%; multidisciplinary reasoning with tools jumps from 54.7% to 57.9%; agentic computer use moves from 82.8% to 83.4%. The only model to complete every case end-to-end on the Super-Agent benchmark, at parity-cost with GPT-5.5. Highest score recorded on the Legal Agent Benchmark. Pricing stays at $5 in / $25 out per million tokens for standard mode; new fast mode runs at 2.5x speed and is 3x cheaper than prior fast variants ($10 in / $50 out per M). Effort dial on claude.ai lets users tune how hard Claude works per task. Ships with Claude Code "Dynamic Workflows" (1,000 parallel subagents per session, research preview on Max/Team/Enterprise).

Bodega One Code ships PreviewInteractionTool. Any vision-capable cloud model (Claude Sonnet 4.6, Opus 4.7, Haiku 4.5, GPT-4o, GPT-5 family, Gemini 2.5 Pro and Flash, Qwen cloud) can screenshot the rendered Preview tab and answer from pixels rather than HTML. llama.cpp vision lands for LLaVA 1.6 Mistral 7B and Moondream2 with paired mmproj download. First local-first AI IDE to ship the preview-screenshot loop end-to-end with cloud and local vision models.

Google announces Gemini 3.5 Flash at I/O 2026, claiming it outperforms Gemini 3.1 Pro on most benchmarks while running about 4x faster than other frontier models. Positioned as the new default for high-throughput coding and agent workflows in the Gemini API. Cloud-only; no open-weights story.

Google ships Antigravity 2.0 desktop and the Antigravity CLI at I/O 2026. The new CLI is available to all users with no waitlist and replaces Gemini CLI, which migrates into Antigravity CLI. New capabilities: specialized sub-agents, cross-platform terminal sandboxing, credential masking, hardened Git policies, and Managed Agents in the Gemini API. Cloud-required, Gemini-only; usage still gated by Google AI Pro and Ultra subscription tiers.

RooCodeInc/Roo-Code repository archived May 15 after hitting 3 million cumulative installs. Co-founder Matt Rubens cites going all-in on Roomote, a Slack-first cloud agent with GitHub, Linear, Jira, Asana, Notion, Sentry, and Vercel integrations. A community fork at Zoo-Code-Org/Zoo-Code (zoocode.dev) carries the VS Code IDE extension forward with an official handoff. Final residual marketplace installs: 1,636,212.

Ollama ships v0.23.3 (May 12) and v0.23.4 (May 14) in quick succession. Adds support for passing local image files as inputs to vision models, and fixes a bug in Claude tool result formatting when the response references a local image path. Update via `ollama update` or download the latest installer.

Cursor 3.4 ships team controls for configuring cloud-agent development environments, including multi-repo support so a single agent can work across every repo it needs with re-use across sessions. Bugbot now offers three effort levels: Default (0.7 bugs per run on average, 79% resolved at merge time), High (0.95 bugs per run, more cost and latency), and Custom (natural-language description of when to use each). Bugbot is now metered usage-based inside Cursor Individual tier; existing customers cut over at next renewal after June 8, 2026.

Cline 2.0 ships as a layered TypeScript stack with @cline/sdk as a pluggable agent harness underneath VS Code, JetBrains, and the CLI. Available as `npm i -g @cline/sdk` or as sub-packages (provider layer, stateless loop). Cline reports 7 million developers across their products migrating to the new harness; CLI and Kanban already on it. Experimental connector channels link agents to Telegram, WhatsApp, and Slack. Same infrastructure-pivot play as OpenCode.

LM Studio 0.4.13 ships MLX engine v1.8.1 with parallel predictions for vision-capable models including Qwen 3.5, Qwen 3.6, and Gemma 4. Performance gains across the board on Apple Silicon. Also bundles a newline-on-paste bug fix plus security hardening. Recommended for all Mac users.

May 7 leaderboard snapshot shows Anthropic's unreleased Claude Mythos Preview leading SWE-bench Verified at 93.9%, ahead of Claude Opus 4.7 at 87.6% and GPT-5.5 at 88.7%. Mythos is not generally available; access is gated through Project Glasswing, an invitation-only program for 12 founding organizations and roughly 40 vetted critical-infrastructure operators.

EU co-legislators reach political agreement on the Digital Omnibus on AI. Stand-alone high-risk AI systems now apply December 2, 2027; embedded high-risk systems August 2, 2028. The transparency grace period for AI-generated content drops from 6 months to 3, with a new deadline of December 2, 2026. The August 2, 2026 GPAI deadline still stands until formal adoption.

Cursor 3.3 surfaces a per-agent breakdown of context window consumption across rules, skills, MCP servers, and sub-agents. The goal is diagnosing context bloat in long agentic sessions and trimming the surfaces that eat the largest token share. Released a week before Cursor 3.4 introduced multi-repo cloud agents.

OpenAI rolls GPT-5.5 Instant out as the new default model in ChatGPT for free users, two weeks after GPT-5.5 and GPT-5.5 Pro shipped to the API on April 23-24. GPT-5.5 leads SWE-bench Verified among general-purpose models at 88.7% per the May 7 leaderboard snapshot, behind only Claude Mythos Preview at 93.9%.

Mistral merges Medium, Magistral, and Devstral 2 into one 128B dense multimodal model with a 256K context, released under a modified MIT open-weights license. Scores 77.6% on SWE-bench Verified. API priced at $1.5 input / $7.5 output per million tokens. Self-hostable on 4 GPUs. Ships alongside Vibe remote agents that run async coding sessions in the cloud.

Novee Security publishes details on CVE-2026-26268, an arbitrary code execution flaw triggered the moment Cursor's AI agent operates on a malicious cloned repository. NVD rates it 9.9 critical; Cursor contests with its own 8.0 high score. Root cause is Git hook execution under autonomous agent control, not a Cursor logic bug per se.

GitHub announces every Copilot plan transitions to token-based billing on June 1, 2026, replacing premium request units (PRUs) with GitHub AI Credits. Credits are consumed by token usage (input, output, and cached) at each model's published API rates. Paid plans include monthly base credits matched to subscription price ($10 Pro, $39 Pro+). A preview invoice ships in early May so users can estimate costs before the cutover. Code completions stay unlimited on paid plans.

DeepSeek open-sources V4-Pro (1.6T total / 49B active) and V4-Flash (284B / 13B active) under MIT license, both with 1M token context. V4-Pro scores 80.6% on SWE-bench Verified per DeepSeek release notes, matching Gemini 3.1 Pro and within a point of Claude Opus 4.6 Max. Hybrid attention (CSA + HCA) cuts FLOPs to 27% of V3.2 at 1M tokens.

The April 24 deadline set by GitHub on March 25 arrives. Copilot Free, Pro, and Pro+ users who did not opt out before today are now contributing interaction data - code context, suggestions, file names, repo structure - to GitHub AI model training by default. Business and Enterprise plans are unaffected. The original announcement received 172 downvotes on GitHub's own forum. Opt out at github.com/settings/copilot/features.

Alibaba releases Qwen3.6-27B, a dense 27B coding-focused model with hybrid Gated DeltaNet attention. Scores 77.2% on SWE-bench Verified, 83.9% on LiveCodeBench v6, and 94.1% on AIME26. New highest score of any locally-runnable model, beating the prior leader Qwen3.6-35B-A3B (73.4%) released six days earlier and matching the closed Qwen3.5-397B-A17B MoE despite having 14x fewer total parameters. Apache 2.0. 262K native context (1M+ via YaRN). Runs on a single RTX 4090 or 24+ GB Mac at Q4_K_M (~17 GB).

OpenCode crossed 146,000 GitHub stars - up from 95K on April 10 - making it the fastest-growing open-source coding agent project tracked. Terminal-native, MIT license, 75+ LLM provider support including local models. Now surfacing in AI search results for "AI coding agent", "open source Cursor", and "local coding agent" queries. Worth monitoring as competitive pressure on commercial tools.

Moonshot AI releases Kimi K2.6, a 1 trillion parameter MoE with 32B active per token. Self-reported 80.2% SWE-bench Verified (Moonshot in-house framework, not canonical harness). Distinctive capability: orchestrates up to 300 sub-agents simultaneously in a single session, with documented 13-hour autonomous coding runs. Requires ~250 GB VRAM. Modified MIT license with a commercial branding clause for deployments over 100M MAU or $20M annual revenue.

GitHub suspends new sign-ups for the Copilot API as the product transitions from flat-rate to per-token billing in Copilot Enterprise. Existing API users are unaffected during the transition window. Follows the March 25 announcement that Copilot training data opt-out defaults would change effective April 24.

Ollama ships two releases in quick succession - v0.21.0 (Apr 17) and v0.21.1 (Apr 21). Flash attention enabled for Gemma 4 on compatible GPUs, giving a meaningful speed boost on RTX 40-series and M-series chips. Bug fixes for Metal backend on macOS. Update via `ollama update`.

OpenAI releases o3 and o4-mini simultaneously. Key shift: both models can invoke tools (web search, code interpreter, image analysis) during the reasoning chain, not just after. o4-mini is the first OpenAI cloud model to clear 70% on SWE-bench Verified. Replaces o3-mini as the cost-efficient reasoning choice. Neither model is open-weight.

Alibaba releases Qwen3.6-35B-A3B, a coding-focused MoE with 35B total parameters and 3B active at inference. Scores 73.4% on SWE-bench Verified. At Q4_K_M quantization (~20 GB) it runs on a single RTX 3090 or 4090. Apache 2.0 license. Ollama support same-day. A strong pick for 24 GB GPU owners who need above-70% coding performance.

Anthropic releases Claude Opus 4.7 via API, Amazon Bedrock, Vertex AI, and Microsoft Foundry. Meaningful improvement over Opus 4.6 on advanced software engineering tasks. Not locally runnable, but relevant for BYOLLM users who route cloud requests through OpenRouter or the direct Anthropic API. Sets a new ceiling for what the cloud-boost tier of a local-first setup can reach.

Mistral releases Devstral 2 Large at 123B parameters, scoring 72.2% on SWE-bench Verified. Follows Devstral Small 2 (68%, March 24). The Large version handles longer context and more complex repository-level tasks. Apache 2.0 license. Requires a multi-GPU setup at any reasonable quantization.

OpenCode, an open-source terminal-native coding agent, surpasses 95,000 GitHub stars and begins surfacing in AI search responses for "AI coding agent" and "open source Cursor alternative" queries. Supports 75+ LLM providers including local models. MIT license. A free, local-compatible alternative to commercial agents gaining real user momentum.

MiniMax releases M2.5, a 229B MoE with 10B active parameters, scoring 80.2% on SWE-bench Verified. The first open-weight model to cross the 80% threshold on the canonical leaderboard. Runs in ~128 GB VRAM with 3-bit quantization. Commercial-OK license.

Z.ai releases GLM-5.1, a 754B / 40B active DSA MoE scoring 58.4% on SWE-bench Pro (a harder evaluation than standard SWE-bench Verified). Claims #1 open-source ranking at release. Novel GLM_MOE_DSA hybrid architecture. Trained entirely on Huawei Ascend chips. Requires ~640 GB VRAM (8x H100 equivalent). MIT license.

Ollama ships same-day support for Llama 4 Scout and Maverick. Run Scout: `ollama run llama4:scout`. Run Maverick: `ollama run llama4:maverick`. Scout Q4_K_M weighs ~67 GB and needs a 48 GB GPU, dual 3090s, or a high-memory Apple Silicon Mac. Maverick is significantly lighter and runs on 24 GB setups.

Meta releases Llama 4 with two models. Scout: 109B total / 17B active (MoE, 16 experts), 10 million token context window - the largest of any open-weight model at launch. Maverick: smaller, faster, competitive on coding benchmarks. Scout SWE-bench: 47.3%. First Llama models with MoE architecture. Both are natively multimodal (text and images). Llama 4 license: commercial use permitted under 700M MAU threshold.

Cursor ships version 3 with a full Agents Window for running parallel agents across local, worktree, and cloud environments. New /best-of-n command runs the same task across multiple models in isolated worktrees and shows results side by side. Full worktree integration lets you run agents without touching your working directory.

Ollama 0.20 ships same day as Gemma 4 with full support for all four Gemma 4 sizes. Run the E2B edge model or the 31B dense frontier model: `ollama run gemma4`. Builds on the MLX backend improvements from 0.19.

Google clarifies the Gemini API free tier: Flash and Flash-Lite models remain free for development. Gemini 3.1 Pro Preview requires a paid key; the free tier shows "Not available" for Pro-class models. The distinction matters for local AI builds that use Gemini as a cloud boost: Flash is free, Pro is not.

Google DeepMind releases Gemma 4 under Apache 2.0, a significant license upgrade from the old restrictive Gemma terms. Four model sizes: E2B (edge/mobile), E4B (~5 GB VRAM), 26B MoE (~14 GB), and 31B dense (~19 GB). The 31B ranks #3 on the Arena AI text leaderboard. AIME reasoning score jumps from 20.8% on Gemma 3 to 89.2%, the largest single-generation reasoning leap in open-source LLMs. Full Ollama support: `ollama run gemma4`.

PrismML releases two 1-bit quantization models under Apache 2.0. Bonsai 8B fits in 1.15 GB of RAM, down from 16+ GB for a standard 8B model. Bonsai 1.7B runs in 240 MB and is confirmed working on Raspberry Pi 5. No Ollama support yet. Load via Hugging Face Transformers or PrismML's llama.cpp fork. First 1-bit models to hold up on real tasks including chat, summarization, and tool calling.

Ollama 0.19 ships an MLX backend for Apple Silicon. Measured gains: 1,810 tokens/s prefill vs 1,154 in v0.18 (1.6x), and 112 tokens/s decode vs 58 in v0.18 (2x). Add `--backend mlx` to any ollama run command to use it. Affects all models on M-series chips.

GitHub changes the default for Copilot Free, Pro, and Pro+ users: interaction data (inputs, outputs, code snippets, and context) will be used to train AI models unless you opt out. Effective April 24, 2026. Business and Enterprise plans are not affected. Opt out at github.com/settings/copilot/features before the deadline.

Windsurf replaces its flexible credit system with fixed daily and weekly quotas. Pro jumps from $15 to $20/month. New Max tier at $200/month offers higher quota ceilings. Tab autocomplete stays unlimited on all plans. Owned by Cognition AI since July 2025.

Augment Code moves to four tiers: Community (free), Indie ($20/mo), Standard ($60/mo), Max ($200/mo). Inline completions and Next Edit will be removed for all non-Enterprise plans on March 31. AI chat and code review continue. The company is pivoting resources toward Intent, its macOS-only agent orchestration product.

Anthropic confirms flat per-token pricing across the full 1M context window for Claude Opus 4.6 and Sonnet 4.6. A 900K-token request is billed at the same rate as a 9K-token request. Prompt caching and batch discounts apply at standard rates throughout.

"Small" refers to 6B active parameters per forward pass. Total model size is 119B across 128 experts with 4 active per token. At Q4_K_M quantization all 119B weights load into ~67 GB VRAM. The payoff: 256K context window, Apache 2.0 license, strong multilingual benchmarks. Runs on Mac Pro 192GB unified memory or a single A100 80GB. No Ollama support at launch. Load via llama.cpp or Hugging Face Transformers.

Cursor expands beyond its $20/mo Pro plan with two new tiers: Pro+ at $60/month (3x usage credits) and Ultra at $200/month (20x usage). Both use a credit-based model where your monthly spend equals your credit pool across OpenAI, Claude, and Gemini.

Mistral releases Devstral Small 2 with 68% SWE-bench Verified, up from 46.8% in v1. The largest single-model improvement of the year so far. Fits in 16GB VRAM at Q4.

Kilo Code ends its free trial period on March 23. KiloClaw Cloud agent is $9/mo ($8/mo on a 6-month plan), down from launch price of $49/mo. AI inference costs are separate. The core VS Code extension remains free and open-source under Apache 2.0 with bring-your-own-key.

Alibaba releases Qwen3.5-27B scoring 72.4% on SWE-bench Verified. Fits comfortably in 24GB at Q4_K_M. Replaces Qwen2.5-Coder-32B as the recommended model for RTX 3090/4090 users.

Zhipu AI releases GLM-4.7-Flash, a Mixture-of-Experts model with 30B total parameters but only 3B active at inference. Runs on 6GB VRAM while performing like a 14B dense model. A solid pick for budget GPUs.

Augment Code drops inline code completions from non-Enterprise plans. The company pivots to Intent, a macOS-only agent orchestration product. $252M in funding, narrowing focus.

Tabnine Enterprise is $39 per user per month. The company positions itself as the "only air-gapped AI platform" for regulated industries. The free tier was removed in April 2025 and the Dev/Pro plan was sunset in October 2025 (release 5.24.0). Tabnine is now enterprise-only.

Microsoft releases Phi-4-mini scoring 60.5% on LiveCodeBench at just 5GB VRAM. The best reasoning-per-watt ratio in its size class. Good fit for laptops and low-VRAM setups.

OpenAI publishes GPT-OSS-20B, their first open-weight model. MoE architecture with 20B total parameters and 3.6B active. Runnable locally on consumer hardware. A notable shift in direction for OpenAI.