# anystrata runs big MoE models on any GPU, BC-250 included

> anystrata, a Vulkan fork of PolyStrata, reruns its BC-250 benchmarks on the current main branch and updates its mascot art.

- Published: 2026-11-05T06:25:25Z
- Score: 7/10
- Tags: bc250, ai, tool, port
- Port / platforms: - -> Linux / Steam Deck (Tool)
- Web page: https://emu-radar.customfw.xyz/news/anystrata-vulkan-moe-llm-bc250/
- Sources: https://github.com/bjaan/anystrata/commit/0c42b0db71f61db01d9fbc360924a26090f683f7

**anystrata** is a fork of PolyStrata, an engine that runs Mixture-of-Experts language models. This version uses Vulkan, so the driver you already have is enough. No vendor SDK is needed, just one `make`.

The idea is simple: the experts stay in VRAM, and each token reads only the few it routes to. The README reports Qwen3.6-35B-A3B at 64K context on a BC-250, a repurposed PS5 board with 16 GB of unified memory that ROCm does not see. All 10,240 routed experts stay resident, and it decodes about 19 tokens per second. The server speaks the OpenAI API, so coding agents can use it.

This commit refreshes the measurements on the current main branch and updates the mascot art (silly.png and any-cyan-approves.png). The engine itself is unchanged. The commit message does not list new performance numbers.

To try it, you need Qwen3.6 GGUF files and a working Mesa RADV setup. The model files and some parts of the code carry their own licenses, so check those before you use them.
