PrismML's Ternary Bonsai 27B Runs Full Reasoning on a Laptop at 7.2 GB

A new low-bit release is challenging the assumption that going below four bits per weight means throwing away most of a model's reasoning ability. Ternary Bonsai 27B from PrismML packs a 27B-parameter model into roughly 7.2 GB using ternary weights, and keeps almost all of the full-precision behavior intact. The model, based on Qwen3.6-27B, targets Apple Silicon and CUDA with custom kernels and ships under Apache 2.0.

The pitch is direct: full 27B-class reasoning in ternary transformer weights, on everyday laptops, roughly 9.4x smaller than FP16 while retaining 95% of FP16 intelligence and hitting about 26 tokens per second on an Apple M5 Pro laptop.

Ternary weights, honestly labeled

Every weight in the language model takes one of three values from the set {-1, 0, +1}, with a single FP16 scale factor shared across each group of 128 weights. A ternary value carries log2(3) or about 1.585 bits of information, so the effective storage cost is roughly 1.71 bits per weight once you amortize the 16-bit scale over 128 weights, giving an idealized 9.4x reduction versus FP16.

The extra zero state (compared to a pure binary sign bit) gives the model a more expressive weight alphabet, which is why this is positioned as the quality-oriented operating point of the family. A 1-bit companion at 1.125 bits per weight and 3.9 GB is sized to fit an iPhone.

Unlike conventional low-bit builds whose advertised labels understate their true average bit-width (a widely-used 2-bit build of Qwen3.6-27B is really 2.8 bits per weight at 9.4 GB), the Bonsai representation carries a bit-width that matches its name. The ternary encoding covers embeddings, attention projections, MLP projections, and the LM head, with no high-precision escape hatches hiding inside layers labeled as low-bit.

Surviving the sub-4-bit cliff

The interesting benchmark story is what happens when other quantization methods drop below 4 bits. Evaluated across 15 benchmarks in thinking mode, the full-precision Qwen3.6-27B averages 85.07. Ternary Bonsai 27B hits 80.49 (94.6% of FP16), higher than the conventional IQ2_XXS build at 72.73, while occupying less than two-thirds of its footprint.