Which model should I run on the on-device AI, and how fast is it really?

[MTSistemi]

The board has an RDNA 2 GPU (gfx1013) with no ROCm support, so the usual accelerated path
is closed. SkillFishOS runs Unsloth Studio on llama.cpp's Vulkan backend instead.

[MTSistemi]

Measured on our board with Qwen3-1.7B: 210 tokens/second on the GPU over Vulkan
against 41 on the CPU alone. Five times faster, and the difference between usable and
not.

Practical advice:

• Up to ~4B parameters, quantised, is the comfortable range. Models come straight from
  the Hugging Face catalogue inside the app.
• Watch VRAM: the split is set in the Tuner (VRAM section, takes a reboot). 8 GB is the
  sensible default for mixed gaming and AI.
• The service listens on loopback only. From another machine you reach it through the
  Remote Manager, which authenticates against your system account — the model server itself
  is never exposed.
• Unsloth releases frequently. sudo skillfish-unsloth-update runs their official installer
  as your user, cleans up the desktop entry it drops, restarts the service and tells you
  which version you went from and to.

If you get numbers on a different model, post them in Show and tell — we only have one
board and one set of measurements.