Strata Local AI: Run Qwen3.8-Flash-Next on a Gaming PC

Short answer: Strata is a free, MIT-licensed program that runs Alibaba's open-weight Qwen3.8-Flash-Next model on a PC with a 12 GB graphics card and 48–64 GB of memory. It keeps the most-used parts of the model on the graphics card and the rest in system memory, and serves it to Claude Code, Codex, OpenClaw and similar tools as if it were a cloud API. The project reports 50–94 tokens a second on an RTX 5070. The software and the tokens are free; the price is the hardware, the electricity and the model's quality compared with paid models.

Strata at a glance

From Strata's README and model guide, read 8 November 2026. Details for each item are below.

What Strata is, and what it isn't

Strata is an inference engine: the software that loads a model and generates text with it. It was built around one model, Qwen3.8-Flash-Next, which Alibaba's Qwen team released as open weights on 26 August 2026 as an experimental preview of its next architecture. The model has 125 billion parameters, plus a separate 51-billion-parameter lookup table, but uses only about 6 billion of them for each word it writes. That design is what makes Strata possible.

Three things it is not:

• Not the Qwen API. Alibaba also sells a hosted model called Qwen3.8 Flash at $0.15 / $0.47 per 1M tokens. Flash-Next is the open-weight preview you download and run yourself.
• Not a general model manager. Strata also runs Qwen3.6-35B-A3B and, slowly, DeepSeek V4.1 Flash, but it is built for one model family. Ollama, LM Studio and llama.cpp run thousands of models.
• Not the other Strata products. Several unrelated companies use the name, including healthcare analytics, identity security and legal software. This guide is about the open-source project on GitHub.

The model's own licence is the Qwen Community License 1.0, not Strata's MIT licence, and it sets conditions on commercial use. Read it before you build a product on the model.

How a 125B model fits on a 12 GB card

The model is split into 24,576 small "experts", and each word needs only 10 of them. Strata spreads the work across your PC:

• Graphics card: the parts every word uses (attention, routing, shared experts), the conversation's memory, and a cache of the experts your conversation uses most. The project puts it at about 700 more experts for every extra gigabyte of graphics memory.
• System memory: all 24,576 experts. Your processor computes the ones that aren't on the graphics card at the same time as the card works on the cached ones.
• SSD: the 28.8 GB lookup table, read a few rows per word.
• Drafting ahead: the model guesses up to three words at once and checks them in one pass. Strata reports 2.4–3.2 words accepted per pass on average, which is where much of the speed comes from.

That is why system memory matters more than the graphics card: the whole model has to sit in RAM. The graphics card mainly decides how fast it runs.

Strata system requirements

Strata's rule of thumb is that the model fits when your RAM is at least the size of its experts plus about 10 GB. The download is 66–76 GB for the three smaller versions, and the SSD also holds the 28.8 GB table.

From Strata's model guide, read 8 November 2026. With 32 GB of RAM, a 24 GB graphics card can run Q2_0, IQ2_XS and the Coder in Strata's low-RAM mode; a 12–16 GB card runs the Coder.

• Graphics card: 12 GB or more. Strata supports NVIDIA RTX 20, 30, 40 and 50 series; AMD RX 7700 to 7900, RX 9060 and 9070 and the Radeon AI PRO R9700 (RX 6800 and 6900 are experimental); Intel Arc on Linux; Apple Silicon M1 or newer; AMD Ryzen AI Max (Strix Halo); and NVIDIA DGX Spark.
• System: Windows or Linux on x86-64, or macOS 15 or newer on Apple Silicon.
• Storage: a fast NVMe SSD with about 100 GB free for the smaller versions (the download plus the 28.8 GB table).

If you're buying hardware for this, memory is the cheap part to get right: 64 GB of RAM opens every version. Our Local AI pages show which other models fit the same machine.

How fast Strata runs

These figures come from Strata's own measurements and from users who posted results to the project. We haven't measured them ourselves, engine versions differ between rows, and speeds change with almost every release.

Sources: Strata's model guide, community benchmarks page and release 0.2.11 notes, read 8 November 2026. "Long context" is the conversation length the speed was measured at.

Reading your prompt is much faster than writing: Strata reports 1,620–2,650 tokens a second for a 32K-token prompt on the RTX 5070, and it keeps the conversation so far between turns, so an agent doesn't pay to re-read it every step. For comparison, 50 tokens a second is faster than most people read.

How good are its answers?

Lower versions trade quality for speed and memory. The project says IQ3_S "matches the full model on the published tests", and lists the smaller versions as good to great. Outside reviews are mixed: Hackaday found it confused facts in general questions but handled code well, and Strata itself warns that the Coder version is weak outside code and in non-English text, and that answers can differ between runs. Treat it as a capable open model running on budget hardware, not a replacement for the strongest paid models on hard work.

How to set up Strata with Claude Code, Codex or OpenClaw

Strata runs a local server that speaks both the OpenAI and the Anthropic API formats, so most coding agents can use it in place of a paid API.

1. Install. On Windows, in PowerShell: irm https://github.com/Niko1221/Strata/releases/latest/download/install.ps1 | iex, then open a new terminal. On Linux or macOS: curl -fsSL https://github.com/Niko1221/Strata/releases/latest/download/install.sh -o install.sh && sh install.sh.
2. Start it with strata launch, choose the version and the context size, and it starts the server at http://127.0.0.1:8080. The default context is 32K tokens; you can choose from 8K to 512K, and each step up costs more memory.
3. Connect your tool with strata connect, which prints the settings for OpenCode, Codex CLI, Claude Code, oh-my-pi, pi, Hermes Agent, OpenClaw and the DeepSeek Harness. strata connect <tool> --check sends one prompt and one tool call to confirm it works.

Things to know: OpenAI-style tools use the server address plus /v1, while Anthropic-style tools such as Claude Code use the address on its own. Claude Code also needs CLAUDE_CODE_MAX_CONTEXT_TOKENS set to match Strata's context, and OpenClaw needs Node 24.16 or newer. By default the model thinks at its highest effort setting, so it writes more tokens before each answer.

Strata is a server on your machine. Keep it on localhost unless you know how to secure it.

Strata vs the API: what it saves

Take the coding agent in our agent cost calculator: 40 steps a task, 20 tasks a day, 22 working days a month. Through an API, that month costs:

The agent calculator's coding agent example: 40 steps a task, 90% of repeated input cached where the model has a cache price, 20 tasks a day for 22 days. List prices checked 13 Oct – 11 Nov 2026; this table updates automatically.

On Strata the tokens are free, so the running cost is electricity. Each of those tasks writes about 15,000 tokens and reads about 100,000 tokens of new text (tool results, files and its own earlier steps). At the speeds above, that's roughly 5 minutes of generation per task, or under 2 hours a day for 20 tasks. If your PC draws 300 W while generating (a plug-in power meter will tell you), 22 working days of that is about 12 kWh: roughly $2.40 a month at $0.20 per kWh.

So the sums are simple:

• Against the cheapest APIs, local saves little. GPT-6 Luna runs that month for about $25.98. A PC with 64 GB of RAM and a 12 GB card costs far more than years of that.
• Against mid-range models, the gap is real: Claude Sonnet 5.5 costs about $427 for the same month. If Qwen3.8-Flash-Next is good enough for your work, hardware you already own pays for itself quickly.
• Privacy and limits are the other reasons: nothing leaves your machine, there are no rate limits or weekly caps, and it works offline.

Try your own hardware and electricity price in the local LLM cost calculator, and see whether a Claude Code plan or an API key is the better comparison for your use.

Strata vs Ollama, LM Studio and llama.cpp

• Strata if you want the largest model your gaming PC can run, mainly for coding agents. It's specialised: few models, and built for one person rather than a team.
• Ollama or LM Studio if you want to try many models with little setup. For models that fit entirely on your graphics card, such as Qwen3.6-35B-A3B on a 24 GB card, they're the simpler choice.
• llama.cpp if you want the most control. Strata itself builds on ggml, the library underneath llama.cpp.

Our Local AI section lists which models fit each graphics card and Mac, with estimated speeds, and Qwen3.5-9B settings for OpenClaw covers a smaller model that runs on almost anything.

Before you rely on it

• It moves fast. The project is about six weeks old and ships releases weekly (0.2.12 came out on 7 November 2026, with a list of known problems). Commands and speeds in this guide may change.
• Speeds are reported, not guaranteed. Your processor, memory speed and SSD matter as much as the graphics card, and running out of RAM slows everything sharply.
• Check the model licence before commercial use.

Questions

Is Strata free?

Yes. Strata is free, open-source software under the MIT licence, and the model is a free download. The model has its own licence, the Qwen Community License 1.0, which sets conditions on commercial use.

Can Strata run with 32 GB of RAM?

Partly. With 32 GB, a 12–16 GB graphics card runs the coding-only version, and a 24 GB card also runs Q2_0 and IQ2_XS in Strata's low-RAM mode. The other versions need 48 or 64 GB.

Does Strata work on a Mac?

Yes, on Apple Silicon (M1 or newer) with macOS 15 or later. It also runs on Windows and Linux PCs.

Is Qwen3.8-Flash-Next as good as Claude or GPT?

We haven't found an independent comparison with paid models yet. Strata says its largest version matches the full model on the published tests, and early reviews found it strong on code but less reliable on general facts. Test it on your own work before you drop a paid model for anything hard.

Sources: Strata on GitHub (README, how-it-works, model guide, community benchmarks and release notes), Hackaday and DataNorth on the model's release. Read 8 November 2026. API prices update automatically.