Introduction

Bonsai 2 is a family of open-weight vision-language models from PrismML, in four sizes from 2B to 27B. Every language-model weight is ternary, one of {−1, 0, +1}, which works out to about 1.7 bits per weight instead of 16. That holds across the whole model, embeddings, attention, MLPs, and the LM head, with no high-precision escape hatches behind the low-bit label, and it is what makes the models small enough to run on laptops, phones, and single GPUs that could never load a full-precision model of the same class. The practical consequence: Bonsai 2 27B fits in 5.95 GB instead of ~54 GB at FP16 and keeps 98.2% of the FP16 model’s benchmark average. Bonsai 2 2B fits in 0.43 GB. All four sizes take images as well as text, carry a 262K-token context, call tools natively, think before they answer, and ship under Apache 2.0. How the weights are packed, and which runtimes read them, is covered in Formats & runtime support.

​

The lineup

Four sizes, from a 27B flagship to a 2B that fits in half a gigabyte: GGUF sizes are the language model alone; add the Q8_0 vision projector (0.36–0.63 GB) for image input. The MLX packs bundle the FP16 vision tower. 27B is the demo repo’s default; pick a smaller size with BONSAI_MODEL. Every size also has a Q2_0 GGUF that runs on upstream llama.cpp (the 27B’s is in a separate repo); see Formats & runtime support. Bonsai Image 4B is a separate family of compact text-to-image models built with the same low-bit approach.

​

Why low-bit models

• They run where full-precision models can’t. Bonsai 2 27B needs 5.95 GB of weights, so it runs on an everyday laptop; the same model at FP16 needs 54 GB and doesn’t fit at all. The 2B’s weights are half a gigabyte.
• They generate faster. Token generation is memory-bandwidth-bound. Moving under two bits per weight instead of 16 means each token needs a fraction of the memory traffic, so decode speed scales up on the same hardware: ~130 tok/s for the 27B on an RTX 5090, ~47 tok/s on an M5 Max laptop.
• They use less energy. The 27B decodes at 27.5 W on an M5 Pro’s GPU rail, and at about 2 J per token on an RTX 5090.
• They keep data local. Inference runs entirely on your machine. Nothing leaves it.

​

Earlier models

Bonsai 2 supersedes the first Bonsai generation: the 1-bit Bonsai and 1.58-bit Ternary-Bonsai families at 27B, 8B, 4B, and 1.7B. Those weights stay available under the prism-ml org on Hugging Face (the 27B ones are in the Bonsai 27B collection), and the demo repo keeps a guide to running them in Bonsai1_README.md. These docs cover Bonsai 2 and Bonsai Image.

​

Where to go next