Model releases  ·  11 August 2026

Bonsai 27B: a 27-billion-parameter model in 3.8 GB

Every weight is one bit. Not quantised down to one bit after the fact — trained that way, which is the difference between a model that survives compression and a model that was never large in the first place.

At a glance

Released byPrism ML
PublishedJuly 2026
Parameters~27.3B binary language weights (~24.8B backbone + ~2.5B embedding and LM head) plus a ~0.46B vision tower
Weight representation1 sign bit per weight; every group of 128 weights shares one FP16 scale
BackboneQwen3.6-27B hybrid attention, approximately 75% linear
Context length262,144 tokens
ModalitiesText and image (separate mmproj projector)
LicenceApache 2.0
SizesQ1_0 3.80 GB · F16 53.81 GB · mmproj 0.63–0.93 GB
Adoption2.7 million downloads on the GGUF repository
Hosted atprism-ml on Hugging Face

Trained at one bit, not compressed to one bit

This is the distinction the whole release rests on. Taking a normally trained model and quantising it to one bit per weight destroys it — there is not enough resolution left to represent what the training produced. Bonsai is trained with that constraint present, so the weights it learns are ones that a single bit can express.

The scheme is simple to state: each weight is a sign bit, where 0 means −scale and 1 means +scale, and every group of 128 weights shares one FP16 scale factor. That works out to roughly 1.125 bits per weight in the packed format.

The result is a 27-billion-parameter model in 3.8 GB — about a fourteen-fold reduction against a conventional half-precision build of the same parameter count.

Long context, made affordable by the backbone

Bonsai carries a 262K-token context. Prism ML attributes its practicality on-device to two things: a Qwen3.6-27B hybrid-attention backbone that is roughly 75% linear attention, and four-bit KV-cache quantisation. Linear attention does not grow quadratically with sequence length, which is what makes a quarter-million tokens tractable outside a datacentre.

What's published

The GGUF repository carries the Q1_0 build at 3.80 GB, an F16 reference at 53.81 GB, a DSpark variant, and separate mmproj vision projectors at 0.63 GB and 0.93 GB. A sibling line, Ternary Bonsai, uses ternary weights instead of binary and is published in GGUF, MLX and AWQ forms.

Adoption is substantial: the Bonsai 27B GGUF repository has passed 2.7 million downloads and the Ternary variant another 839,000.

Runtime support

RuntimeSupportedNotes
llama.cppYesGGUF builds published directly, including the Q1_0 format the model is designed around.
MLXYesPrism ML publishes 1-bit and 2-bit MLX builds for the Bonsai and Ternary Bonsai lines.
Vision inputPartialSupported through a separate mmproj projector that must be loaded alongside the main file.
AWQYesFour-bit AWQ builds published for both lines.

What we don't know yet

Sources

Frequently asked questions

How can a 27B model be 3.8 GB?

Each weight is stored as a single sign bit, with one shared FP16 scale factor per group of 128 weights. That is about 1.125 bits per weight, roughly fourteen times smaller than a half-precision build.

Is Bonsai just a heavily quantised model?

No. It is trained at one bit per weight rather than compressed afterwards. Quantising a normally trained model to one bit does not work; training under the constraint is what makes it viable.

What licence is Bonsai 27B under?

Apache 2.0.

Does Bonsai support images?

Yes, through a separate mmproj vision projector that is loaded alongside the main model file.

What is Ternary Bonsai?

A sibling line using ternary rather than binary weights, published in GGUF, MLX and AWQ formats.

OnDevice LLM is a private AI assistant that runs entirely on your iPhone — no account, no cloud, and nothing you type leaves the device.

Published 11 August 2026