How to run Qwen3.8-Flash-Next on a gaming PC with Strata

On October 4, a Hacker News thread titled “Run Qwen 3.8 Flash Next (125B) on consumer hardware (RTX 4090) at 100T/s” hit 922 points. People wanted to know if it was real. Strata's own measurements back most of it, with conditions. This is the hands-on version: what you need, which size to pick, what the speeds look like, how to install it, and where it falls short.

What is Qwen3.8-Flash-Next?

Qwen3.8-Flash-Next is an open-weights mixture-of-experts model from Qwen with 125B parameters and 6B active per token, plus a 51B n-gram embedding and a 4B draft layer. Qwen calls it an experimental preview of the architecture that will underpin Qwen4. It uses 512 experts, 10 routed plus 1 shared per token, across 48 layers, with a hybrid of Gated DeltaNet and Qwen Sparse Attention. It also takes image input and has a 262,144-token native context, extensible to 1,000,000. The license is the Qwen Community License 1.0, which we cover in the caveats.

The coding numbers are why people care. These are Qwen-reported, at full precision, compared with Qwen3.8-27B:

The gap on SWE-bench Pro is small; the bigger gaps are on DeepSWE and SWE-bench Multilingual. These are vendor numbers, and none of them were run at the 2-3 bit sizes you will actually use here.

What does Strata do?

Strata spreads the model across every part of your PC instead of forcing it all into VRAM, which is how a model this size fits outside a server. It is an MIT-licensed project by Niko1221 that runs one model, Qwen3.8-Flash-Next, in several sizes. The split works like this:

• GPU. Keeps the most-used few thousand of the model's 24,576 experts.
• RAM. Holds all of them.
• CPU. Computes the experts that are not on the GPU, in parallel.
• SSD. Stores a 29 GB lookup table, the 51B n-gram embedding.

On top of that it uses speculative decoding with the model's built-in MTP draft layer. Strata says that gives the same answer 1.6 to 1.8 times sooner. Long prompts are read in chunks of up to 8,192 tokens.

What hardware do you need?

You need 32 GB of RAM or more, a 12 GB+ GPU, and about 80 GB of free disk, and RAM is the number that decides what you can run. Strata lists NVIDIA RTX 20/30/40/50 cards or listed AMD Radeon cards with 12 GB or more, Windows 10/11 or Linux, and a current driver. An SSD is recommended. macOS is not listed.

This is why our GPU guide and the VRAM calculator only get you halfway here. For this model, check your RAM first. Strata's guidance by RAM:

Which size should you pick?

Strata recommends IQ2_XS for 64 GB of RAM, and the Coder version for 32 GB. The installer picks a size for your machine, but you should know what you are choosing between. The RAM-plus-VRAM column is the combined memory the size needs.

The three smaller sizes are a 66 to 76 GB download, plus a roughly 6 GB MTP draft layer. Budget the disk.

The Coder version comes from ISTA-DASLab. It keeps 256 of the 512 experts, chosen on code data, which is how it fits in 32 GB of RAM. Its authors report 91% of the full model's SWE-bench Verified score and 99% of LiveCodeBench. The tradeoff is that it is weaker outside code and in non-English languages. A Strata GitHub issue (#438) describes CJK answers looping. If you only prompt in English, that may not matter. If you do not, it will.

How fast is it?

On Strata's own measurements, a 12 GB RTX 5070 with 64 GB of RAM runs 43 to 94 tokens per second depending on size and context. These are Strata's numbers on its own test rigs, not ours, in tokens per second.

On an AMD RX 9070 XT 16 GB with 47 GB of RAM, Strata measured Q2_0 at 60 tokens per second.

Two numbers you will see quoted elsewhere. Strata estimates an RTX 3090 24 GB at about 100 to 140 tokens per second, and that is an estimate, not a measurement. The “RTX 4090 at 100T/s” in the HN title comes from community reports, not from Strata's own table.

Prompt length matters for agent work. Strata's README budgets about a minute per 30K tokens for the first message of a chat, which is read in full; its Q2_0 benchmark on the 5070 is faster than that. Follow-ups start in seconds.

How do you install Strata?

Download or clone the repo, run the start script for your OS, and let the installer pick a size. Check the repo for current instructions before you start, since the project is only weeks old.

1. Make sure you have about 80 GB free and a current GPU driver.
2. Get Strata from github.com/Niko1221/Strata.
3. On Windows, run START-HERE.bat. On Linux, run ./setup.sh.
4. Let the installer choose a size for your RAM, or override it using the tables above.
5. Wait for the download, then open the app at http://127.0.0.1:8080. It has Chat, Monitor and About tabs.

Expect the first start to be rough. Strata's README says your PC can be slow or stop responding for 1 to 3 minutes while it loads 35 to 55 GB into RAM. That is normal. Do not kill it.

How do you connect a coding tool?

Point your tool at Strata's local endpoint. It speaks the OpenAI API, the Anthropic API, and the OpenAI Responses API. Any API key works, and so does any model name. Reasoning effort can be set to off, low, medium or high.

For Claude Code, set the base URL and start it as usual:

# bash
export ANTHROPIC_BASE_URL=http://127.0.0.1:8080
claude

# PowerShell
$env:ANTHROPIC_BASE_URL="http://127.0.0.1:8080"
claude

For Bodega One Code, add a custom OpenAI-compatible provider. Our app takes any OpenAI-compatible endpoint, on top of its 10+ LLM provider presets (see BYOLLM). Set the base URL to http://127.0.0.1:8080/v1, use any value for the API key, and type any model name. One thing to know: our app does not ship Flash-Next in its built-in model catalog, because the smallest file is 68 GB and runtime support is still settling. Strata is one third-party way to run it locally today.

We have not benchmarked Bodega One Code with Strata. It uses the same OpenAI-compatible endpoint any client uses, and that is all we are claiming. With a local model and air-gap mode, nothing leaves the machine. For a general setup walkthrough, see how to run a local LLM for coding.

If you want to reach Strata from another machine, rerun setup with the host and key flags, and always set a key. On Windows (the Linux steps are in Strata's install guide):

START-HERE.bat --setup --host 0.0.0.0 --api-key <secret>

What are the honest caveats?

It makes a big model runnable on a gaming PC, with real limits. Know these before you spend an evening on it.

• Quality at 2-3 bit. Qwen's benchmark numbers are full precision. Strata's sizes are quantized to 2 or 3 bits. Strata says IQ3_S matches the full model on the published tests, but the smaller sizes and the Coder version give something up. Test on your own code.
• First-start freeze. 1 to 3 minutes of a sluggish PC while it loads into RAM.
• One request at a time. The default is a single request. There is a "parallel": 2 option, but it is slower per answer on a 12 GB card. Multi-agent workflows will queue.
• No macOS. Windows and Linux only, as far as Strata lists.
• Third-party and young. Strata is not our project. The repo was created on September 24, and we can only vouch for what its public docs say.
• The license has conditions. Strata is MIT. The model is under the Qwen Community License 1.0, which is not Apache or MIT. It allows use, modification, distribution, sale, hosting and fine-tuning, with two conditions. Products with over 100M monthly active users or over US$20M monthly revenue must prominently display the model name. And a licensee running a “Model as a Service” or “AI Work Assistant” business, meaning an independent AI product mainly for AI-assisted coding or office productivity, needs a separate license from Qwen before commercial use. Internal use is exempt as long as the model, its outputs or its capabilities are not made available to third parties. In plain terms, using it for your own coding looks fine, and building a coding-assistant product on it needs Qwen's say-so. Read the license, not this paragraph. This is not legal advice.

When is the simpler choice better?

If you have a single 24 GB card and under 64 GB of RAM, run Qwen3.8-27B instead. At about 17 GB in Q4, it fits entirely in VRAM, loads quickly, runs in any standard local runner, and skips the first-start freeze, the one-request limit and the 80 GB download. On SWE-bench Pro, Qwen reports 61.7 for it against 62.5 for Flash-Next, both at full precision. Here you would be comparing a Q4 27B against a 2-3 bit Flash-Next, so that small gap may close or flip. For most day-to-day coding, the 27B is the better trade.

Flash-Next earns the setup when you already have the RAM and want Qwen's larger model, especially on DeepSWE and SWE-bench Multilingual (many programming languages), where the Qwen-reported gap is widest. For a ranked list of what runs on which hardware, see local LLMs for coding. And if you want to try the whole thing inside a local-first IDE, Bodega One Code is free for everyone during the open beta.

Sources

• Qwen3.8-Flash-Next model card, Hugging Face
• Strata repository, GitHub
• Hacker News thread: Run Qwen 3.8 Flash Next on consumer hardware