BULaMU runs 4.8 tok/s on Fire HD 10

This is a small but telling edge-AI demo: tiny, language-specific LLMs can be practical on commodity tablets if you keep the model compact enough. It is more a proof of feasibility than a universal performance claim, but it points in a useful direction for on-device assistants in low-resource languages.

• –The result shows that a 20M model can deliver interactive-ish speed on 3GB RAM hardware, which matters for offline and privacy-preserving use cases.
• –BULaMU’s bigger significance is linguistic coverage: Luganda gets a native model instead of being an afterthought in English-first stacks.
• –Because this is a self-reported single-device benchmark, it should be read as a feasibility demo, not a standardized comparison against other runtimes or quantization schemes.
• –The project’s Hugging Face repo also exposes training scripts and multiple model sizes, which makes it more useful than a one-off benchmark screenshot.