Thesis
Bonsai's reported one-bit model is better understood as a trained low-bit network than as a conventional quantized copy of its source model.
Observed facts
The article says 27 billion parameters would occupy about 4 GB with one bit per weight plus a 16-bit scale per 128-weight group, versus about 54 GB in FP16, excluding the vocabulary table.
1The author's layerwise solver produced much worse perplexity than Bonsai and had layer errors two to three times higher on unseen text than on calibration text.
1The author reports a 70% match in most likely next token between Bonsai and Qwen in the comparisons, despite low hidden-state cosine similarity.
1
Mechanism
The author infers from weight and output comparisons that the block weights were trained in a one-bit representation to preserve model outputs, rather than produced by conventional quantization that keeps weights close to the original matrices.
Why now
Public formats and kernels make one-bit deployment tangible, while the analysis indicates that a storage format alone does not explain a usable model.
Implications
Low-bit deployment may require training methods designed for the representation, not only a more aggressive quantizer.
Evidence boundary
This is an independent analysis of weights and outputs; the author did not inspect PrismML's code, so the training method is inferred, not confirmed.
Watch conditions
Whether PrismML or an independent reproduction publishes training details and out-of-distribution evaluations.