VecSzn/PolyStrata

What it solves

Large AI models usually need a server with hundreds of gigabytes of graphics memory. PolyStrata lets you run them on a normal PC — a laptop with an 8 GB graphics card or a desktop with 32 GB — by sharing the model across your graphics card and system RAM. That means you can chat, write code, and use your own apps and coding agents entirely on your machine, with nothing leaving your PC.

How it works

PolyStrata includes a custom engine that splits the model’s work between your graphics card and RAM. The parts you use often stay on the graphics card, and the rest waits in RAM — like keeping the things you use all the time on the counter and the rest in the pantry. The installer sets up the right engine for your card, downloads the model, and starts a local server. You can then talk to the model in a browser at http://127.0.0.1:8080, or connect any app that supports OpenAI- or Anthropic-style APIs. Ten models are available, most of them experimental, and several can also read pictures.

Who it’s for

People who want to run large AI models privately on their own hardware — developers using coding assistants like Claude Code or Cursor, hobbyists with a capable gaming PC, and anyone who wants to keep their data on-device. It works on Windows, Linux, and Apple Silicon Macs (for one model), and the installer includes a benchmark so you can measure speed on your own machine.

Highlights

• Runs ten large AI models locally, including Qwen3.6, GLM 5.3 Flash, DeepSeek V4 Flash, and more.
• Fast prompt reading — on a laptop it read prompts 2.7–3.5× faster than llama.cpp.
• Works with existing apps via OpenAI- and Anthropic-compatible APIs (e.g., Claude Code, Codex).
• No data leaves your PC; all inference happens on your hardware.
• Free and open source.