Breaking

PrismML Open-Sources Bonsai 27B, a 1-Bit Model That Squeezes 27 Billion Parameters Into Under 4GB

July 19, 2026 at 11:45 EDT

Foundation Models
Open Source
Infra & Chips

PrismML has released Bonsai 27B, an Apache 2.0-licensed multimodal model that compresses 27 billion parameters into roughly 3.9GB using 1-bit quantization, small enough to run locally on a modern smartphone or laptop without any cloud connection. Announced on July 14, 2026, the model is offered in two variants and is available through PrismML's official channels and Hugging Face.

July 14, 2026 · PrismML

A 27-Billion-Parameter Model That Fits in Your Pocket

Bonsai 27B compresses a full multimodal model to roughly 3.9GB using 1-bit quantization — small enough to run entirely on a modern smartphone, with no cloud connection.

3.9GB

1-bit build size (~1.125 bits/weight)

262K

Token context window

27B

Parameters, running on-device

The compression story: same model, a fraction of the footprint

Memory needed to hold the 27B model — column height is proportional to size in GB.

54GB

FP16
full precision

18GB

4-bit
still off phones

5.9GB

Ternary
Bonsai

3.9GB

1-bit
Bonsai

Roughly a 14× smaller footprint than FP16 — the leap that moves a 27B model onto a phone.

Accuracy retained vs. the full model

Average across 15 benchmarks in thinking mode (full model scores 85.0).

95%

Ternary
(80.5 / 85.0)

90%

1-bit
(76.1 / 85.0)

Math and coding hold up well; agentic, tool-calling and vision tasks show larger drops.

Where supporters see wins

Big VRAM savings for agentic work
Ternary GGUF runs on 16GB VRAM
Private, offline on-device agents
Local runs via MLX & llama.cpp

Where critics push back

Qwen3.5-9B at 4-bit may compete
Weak on data analytics & browser tasks
1-bit sometimes trails on coding/math
Calls for independent benchmarking

On-device throughput

11 tok/s

iPhone 17 Pro Max (1-bit)

87 tok/s

M5 Max

163 tok/s

RTX 5090

Apache 2.0 · built on Qwen3.6 27B · text + image input · runs on MLX, CUDA, llama.cpp and WebGPU — pushing the frontier of how large a model can be while still fitting on a phone.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.