# BC-250 Linux Box Gets a Local AI Agent Stack With Qwen 3.5

> A new build script kit brings llama.cpp and Qwen 3.5-9B to the AMD BC-250 mining board running CachyOS.

- Published: 2026-11-03T20:18:53Z
- Score: 6/10
- Tags: bc250, ai, tool
- Port / platforms: Linux / Steam Deck -> Linux / Steam Deck (Tool)
- Web page: https://emu-radar.customfw.xyz/news/bc250-llama-cpp-qwen/
- Sources: https://github.com/nielsek/bc250-llama.cpp/commit/cf67f545f1acabbe3a9d315210bcf185609a7535

Developer nielsek published **bc250-llama.cpp**, a script collection that builds and runs [llama.cpp](https://github.com/ggml-org/llama.cpp) with Vulkan on the AMD BC-250, an ex-mining board packing a cut-down PS5 APU, known in the Linux scene as Cyan Skillfish. It runs [CachyOS](https://cachyos.org/).

The setup serves `unsloth/Qwen3.5-9B-GGUF:Q6_K` headless via `llama-server` on port 8000 for coding tools like [opencode](https://opencode.ai). Because the BC-250 shares 16 GB of GDDR6 between CPU and GPU, the scripts rely on the `ttm.pages_limit=3145728` kernel parameter from the bc250-toolkit to let the graphics driver claim up to 12 GiB of system memory. The setup assumes a fully unlocked BC-250 with all CPU cores and GPU CUs enabled.

The repository has four shell scripts:

1. `./build.sh` installs dependencies, clones, and builds llama.cpp.
2. `./test.sh` verifies GPU detection and checks that the model runs fast enough.
3. `./run.sh` launches `llama-server` inside a tmux session and leaves the desktop to free about 3 GB of RAM.
4. `./stop.sh` shuts down the server and returns to the desktop.

The 9B Qwen model fits the 12 GiB GPU allowance with a 128K context. On this hardware it reads prompts at roughly 400 tokens per second and generates at about 40 tokens per second.
