Held-out vs LFM2-350M (Mac CPU, 60 samples)
Disk24 vs 219 MB · 10×
RAM peak391 vs 526 MB · 1.3×
TTFT p505.5 vs 124 ms · 22×
intent (logprob)1.00 vs 0.42
tool_match0.72 vs 0.47
json_validity0.93 vs 0.98
What this is
A 35 M-param domain-specialised LM (24 MB Q4) running in your browser.
Pretrained on 64 MB EN / MS / ID Wikipedia + system logs; LoRA-tuned
(1.18 M params, 5000 steps) on synthetic factory-ops dialogues.
Good at: EN↔MS / EN↔ID translation,
Manglish replies, maintenance intent labels, short factory-ops Q&A.
Not good at: open chitchat, world
knowledge, math, code. Outputs on those prompts will look broken —
that’s the cost of fitting an LM in 24 MB.