What's New in Local AI

Models, benchmarks, pricing, tools.

Everything that matters for local AI developers. One feed. Updated weekly.

July 2026

3 updates

June 2026

34 updates

May 2026

13 updates

April 2026

23 updates

March 2026

16 updates

Monthly digest with the most important model releases and local AI changes. No spam. Unsubscribe anytime.

Subscribe to updates

Related resources

Local LLM Rankings

See the full rankings with benchmarks

VRAM Calculator

Check if new models fit your GPU

Quick Start Quiz

New to local AI? Start here

← Back to Resources

What's New in Local AI

Updates timeline

July 2026

Anthropic restores Claude Fable 5 worldwide after US export controls are lifted

Anthropic launches Claude Sonnet 5 with a 1M-token context window

Bodega One Code v1.0.0-beta.31.7: Claude Sonnet 5 and Fable 5 join the model list

June 2026

Bodega One Code v1.0.0-beta.31.6: Bodega Mixture (Mixture-of-Agents) and dev-server-free preview

llama.cpp b9840: DeepSeek V4 conversion tooling lands, so you can run V4 weights locally

LM Studio 0.4.17 to 0.4.18: GGUF prompt-template overrides and AMD Strix Halo support

Ollama 0.30.11: unified speculative decoding and a Windows hybrid-GPU fix

Liquid AI LFM2.5-230M: a 230M open-weight model built to run on a phone

Sakana Fugu: a Mixture-of-Agents system delivered as one closed API

Continue.dev is acquired by Cursor and the repo goes read-only

Gemini CLI shutdown gets hard dates: consumer access ends June 18, gone July 17

Bodega One Code v1.0.0-beta.29.x: refreshed local-model catalog, b9670 llama-server, fully air-gapped editor

GLM-5.2: open-weight successor to GLM-5.1, now ahead of GPT-5.5 on SWE-bench Pro

Ollama 0.30.8 to 0.30.10: better KV-cache reuse for long local agent runs

Bodega One Code v1.0.0-beta.29: model routing rules, Verified Private Automation, a smarter agentic loop

Anthropic disables Claude Fable 5 and Mythos 5 worldwide after a US export-control order

Moonshot ships Kimi K2.7-Code: a coding-first 1T MoE tuned for token efficiency

Amazon Kiro adds GitLab support, browser Specs, a $100/mo Pro Max tier, and CLI 2.7.0

Bodega One Code v1.0.0-beta.28: Run Tasks, Runs Inbox, one-click rules import from other editors

Cursor speeds up Bugbot 3x with Composer 2.5 and adds inline /review commands

Zed v1.6.3 adds agent Fast mode and per-command terminal sandboxing

Bodega One Code v1.0.0-beta.27: scheduled agents (Loops), QEL execution proofs, 2x faster verification

Claude Fable 5 is generally available: 1M context, but 30-day retention and no zero-data-retention option

Cohere open-sources North Mini Code 1.0: a 30B-A3B coding model that runs on one GPU

Bodega One Code v1.0.0-beta.26: ACP server, custom agents, Map as a callable tool

Bodega One Code v1.0.0-beta.25: Fleet Parallel, ACP, and a queryable codebase wiki

Cursor June updates: Design Mode, SDK custom tools, Enterprise Orgs, Teams pricing

GitHub adds enterprise-managed Copilot plugins and a sign-in-free BYOK path in VS Code 1.122

NVIDIA releases Nemotron 3 Ultra: a 550B open-weight MoE with open data and recipes

Google releases Gemma 4 12B: a multimodal open model that runs in 16GB

Devin Desktop launches: Windsurf rebrand, ACP support, Devin Local in Rust

Claude Opus 4.8 Dynamic Workflows: 1,000 parallel subagents per session, full docs

GitHub Copilot week-one bills land: 1,180-credit single sessions, devs vow to leave

Microsoft ships its first in-house coding model, MAI-Code-1-Flash, inside Copilot

MiniMax M3: first open-weight frontier coding model with 1M context + multimodal

Bodega One Code v1.0.0-beta.24: local two-model vision on llama.cpp

GitHub Copilot usage-based billing goes live; new Max plan launches at $100/mo

May 2026

Claude Opus 4.8: agentic coding 69.2%, Dynamic Workflows, 3x cheaper fast mode

Bodega One Code v1.0.0-beta.21: agent vision via rendered preview screenshot

Gemini 3.5 Flash: faster than 3.1 Pro, frontier-level at speed

Antigravity 2.0 + Antigravity CLI: Google opens its agentic IDE to everyone

Roo Code archived at v3.54.0: team pivots to Roomote

Ollama 0.23.3 + 0.23.4: vision model image inputs, Claude tool result fixes

Cursor 3.4: multi-repo cloud agents + Bugbot Effort Levels

Cline SDK 2.0: open-source TypeScript agent runtime for any team

LM Studio 0.4.13: MLX engine v1.8.1, parallel predictions for vision models

Claude Mythos Preview tops SWE-bench Verified at 93.9%

EU Council and Parliament agree to delay high-risk AI Act rules

Cursor 3.3: per-agent context usage breakdown

GPT-5.5 Instant becomes default ChatGPT model for free tier

April 2026

Mistral Medium 3.5: 128B dense, 77.6% SWE-bench Verified, modified MIT

CVE-2026-26268: Cursor Git hook RCE rated 9.9 by NVD

GitHub Copilot moves all plans to usage-based billing on June 1

DeepSeek V4 Preview: 1.6T MoE, 1M context, 80.6% SWE-bench Verified

GitHub Copilot training opt-out deadline passes - data collection now active

Qwen3.6-27B: 77.2% SWE-bench Verified at 27B dense, new local SOTA

OpenCode surges to 146K GitHub stars in under two months

Kimi K2.6: 1T MoE, 300 sub-agents, 13-hour autonomous coding runs

GitHub suspends Copilot API sign-ups, shifts to token billing

Ollama 0.21: flash attention for Gemma 4, macOS Metal fixes

OpenAI o3 + o4-mini: reasoning models that use tools mid-thought

Qwen3.6-35B-A3B: 73.4% SWE-bench, fits a single 24 GB GPU

Claude Opus 4.7: Anthropic's strongest coding model to date

Devstral 2 Large: 123B coding model, 72.2% SWE-bench

OpenCode hits 95K stars, enters AI search results for coding agents

MiniMax M2.5 breaks 80% SWE-bench barrier for open-weight models

GLM-5.1: #1 open-source on SWE-bench Pro at release

Ollama ships day-one Llama 4 support

Llama 4 Scout + Maverick: Meta's first MoE family

Cursor 3: agent orchestrator, /best-of-n, worktree support

Ollama 0.20 adds Gemma 4 support

Google Gemini API free tier limited to Flash-family models

Gemma 4: Apache 2.0, #3 open model, runs on a single GPU

March 2026