Bonsai 2 27B

Squeezing a 27-billion-parameter model onto an ordinary laptop is the aim of Bonsai 2 27B, a compressed language model released by PrismML on 17 September 2026. It is built from Qwen3.8 27B, and the company describes the result as small enough to run locally on modest hardware.

How the weights are stored

The saving comes from the way the numbers are held. The model uses ternary weights, each restricted to minus one, zero or plus one, with group-wise scaling kept in 16-bit floating point. PrismML puts the effective figure at 1.76 bits per weight, and the whole model at 5.9 GB, a size that PrismML says is over nine times below that of the full-precision model it was derived from. Despite the compression it keeps a context window of 262,144 tokens and accepts both text and images as input.

Licence and availability

PrismML publishes the weights free of charge under the Apache 2.0 licence, through Hugging Face and GitHub, so they can be downloaded and used without a commercial agreement.

What the vendor says about quality

According to PrismML, the model retains 98.2 per cent of the full-precision model's aggregate benchmark performance, compared with 95 per cent for the first Bonsai 27B released two months earlier. It gives the compressed model an overall 83.9 against 85.4 for the baseline. The seven category figures behind this are PrismML's own measurements, and each is an average over a group of tests the company chose, not a single benchmark result:

• knowledge and reasoning: 83.95, across MMLU-Redux, GPQA Diamond and AA-LCR
• coding: 81.58, across HumanEval+, LiveCodeBench v6, MBPP+ and BigCodeBench
• maths: 96.57, across AIME 2026, AIME 2025, GSM8K and MATH-500
• instruction following: 82.66, across IFBench and IFEval
• vision: 78.59, across CharXiv, A-OKVQA, OmniDocBench v1.6, RealWorldQA and OCRBench v2
• agentic and tool-calling work: 77.57, across tau-squared bench and BFCLv3

Where the compression costs

Each of these figures blends several tests, PrismML has not published how, and no other model in this directory is measured that way, so they appear here only as the vendor's claims and are not stored as scores. The figures do show where the trade-off falls. The biggest losses from compression are in knowledge and reasoning (a drop of 2.71 points against the baseline) and in vision (a drop of 3.05), whereas instruction following comes out slightly ahead of the uncompressed baseline.

Bonsai 2 27B is a compressed model from PrismML, built from the Qwen3.8 27B model and released on 17 September 2026. Its purpose is the footprint: the weights are ternary, limited to minus one, zero and plus one with group-wise scaling, which brings the whole model to 5.9 GB. The model keeps a long context window and takes text and images as input.

What Bonsai 2 27B does

• Compact weights: ternary weights at an effective 1.76 bits per weight, against the full-precision model it is made from.
• Long context: a window of 262,144 tokens.
• Text and image input: accepts text and images as input.
• Open licence: weights are published under the Apache 2.0 licence on Hugging Face and GitHub.

Key facts

• Base model: Qwen3.8 27B.
• Released by: PrismML, on 17 September 2026.
• Benchmarks: the company reports its own composite scores, which are not recorded on this profile because the method is not published.