PrismML's 1-Bit Bonsai 27B Compresses a 27-Billion-Parameter Model to 3.9 Gigabytes

PrismML has released Bonsai 27B, a set of extreme low-bit builds of a Qwen3.6-class 27-billion-parameter model that the company says reduces a model of that size to a size small enough to fit on a phone. According to MarkTechPost, the ternary variant retains 94.6 percent of the 16-bit floating-point (FP16) baseline at 5.9 gigabytes, and the 1-bit variant retains 89.5 percent at 3.9 gigabytes, down from roughly 54 gigabytes at full precision.

The technical claim that matters is the training method. Rather than compressing a finished model with post-training quantization, which typically degrades quality sharply below 4 bits, PrismML says it trained Bonsai as a native low-bit architecture using quantization-aware training. On the reported numbers, the baseline scores 95.3 on a math suite against 93.4 for ternary and 91.7 for 1-bit, and 88.7 on coding against 86.0 and 81.9. The coding gap is where 1-bit compression most reduces accuracy.

These are vendor figures, not independent reproductions. The relevant check is that the weights are open. PrismML has published the models under Apache 2.0 on Hugging Face with llama.cpp and MLX builds, so any lab can rerun the benchmarks and measure the coding regression directly. The superlative claim, that this is the first 27B-class model to run on a phone, is the company's own and has not been independently confirmed.