Strata: A 125B AI Model on Your Gaming PC, Free and Private

If you have a decent gaming PC, you now have a front-row ticket to something genuinely exciting.

An independent developer just shipped Strata, an open-source inference engine that runs Qwen3.8-Flash-Next, a 125 billion-parameter AI model, entirely on your own hardware. Version 0.1.40 dropped this week, and the project is moving fast: it crossed 15,000 GitHub stars in days.

What's happening

"Local AI" used to mean small models: quick and cheap to run, but noticeably less capable than the cloud tools you were paying for. Strata changes that framing. Qwen3.8-Flash-Next is a frontier-class model, 125 billion parameters, in the same weight class as the best cloud AI your team uses today.

Strata spreads the model across your GPU, system RAM, and NVMe drive using a mixture-of-experts architecture. Not every part of the model activates on every query, so a single consumer graphics card can keep up. Community benchmarks in the repo show 60 to 94 tokens per second on mid-range hardware. That is faster than you can read. Open Source For You covered the release this week after it landed.

Once running, Strata serves an OpenAI-compatible API on localhost port 8080. Any tool that already calls an AI (your scripts, agents, or chat apps) can switch to Strata with zero code changes and zero API cost.

Try this this week

• Check your hardware: NVIDIA or AMD GPU with at least 12GB of VRAM, 32GB+ of system RAM (64GB gets the fastest speeds), Windows or Linux.
• Grab the installer from github.com/Niko1221/Strata. One click, no manual setup.
• Plan for the first model download: roughly 80GB of weights land on your NVMe drive.
• Point any AI client or script at http://localhost:8080 and use it exactly like a cloud API.
• Run one real task you are currently paying per-token for: drafting, summarizing, classifying.

The bigger picture

Small teams have been renting intelligence they could own. Strata is the first tool that makes "own it" genuinely competitive with "rent it." Private and cheap was the old pitch for local AI. Now it is also fast and frontier-grade. If you already have a gaming rig, this is a serious option worth trying this week. If you are weighing new hardware for the office, it is one more reason to size up the GPU. The smallest working thing is sometimes the one that costs nothing per query.