IT EN
Usage

On-device AI

The Ollama + OpenWebUI stack accelerated in Vulkan on the BC-250's GPU.

SkillFishOS includes a local AI stack: chat and coding models that run entirely on the BC-250’s GPU, with no cloud and without sending data anywhere. It turns on and off with one click, so you free up GPU and RAM when you want to play.

SkillFishOS AI panel — turns the local LLM engine (Qwen3 14B, Vulkan) on/off and opens web chat, Dockge or OpenCode

Why Vulkan and not ROCm

AMD’s “official” compute stack is ROCm, but it does not support the BC-250’s gfx1013. SkillFishOS therefore uses the Vulkan backend of Ollama, with the Mesa drivers: it works well on the integrated GPU, leveraging the shared memory (and the extended GTT, see GPU).

The components

ComponentRole
Ollama (Vulkan backend)runs the LLM models on the GPU
OpenWebUIweb chat interface (with web search)
DockgeDocker stack management via web

The stack runs in Docker containers with a custom image (Ollama + Mesa’s Vulkan drivers). It is configured not to start on its own (restart: "no"), so it doesn’t steal the GPU from games: you activate it when needed.

The practical reference model is qwen3:14b: it runs 100% on the GPU (~10.7 GB) with the KV cache in f16.

⚠️ On this hardware (RADV driver) the q4_0 quantization of the KV cache corrupts the output: use the f16 cache.

Turning it on/off

A dedicated AI panel (native app, see Native apps) turns the whole stack on and off with one click. Keep in mind:

  • AI and games/Android should not be used together: they share the same GPU and memory;
  • with the stack off, the GPU and RAM are fully available again for gaming.

Sources