Vision on 1-bit LLMs, and a Silent Training Failure

F-03 · measurement1-bit VLMs on edge hardware.

Question
Can vision be grafted onto extreme-quantized, 1-bit language models?
Method
Vision-grafting pipeline onto BitNet and Falcon3 1-bit LLMs, measured across a 6-pipeline VLM benchmark (Qwen2.5-VL, InternVL2).
Result
A boundary-supervision failure that produced no error and no signal at train time and degenerate output at inference, diagnosed and fixed with one line. Measured a CPU/GPU efficiency inversion: ~3× CPU speedup and ~6× lower memory against 4-bit baselines.
Implication
1-bit VLMs are natively suited to CPU and ARM edge hardware.
speed ~3× memory ÷6
1-bit against a 4-bit baseline on CPU. Grey is the 4-bit baseline, blue is 1-bit.