Bonsai 27B is PrismML's family of binary and ternary low-bit versions of Qwen3.6 27B. It retains the base model's vision, reasoning, and tool-use capabilities while making 27B-class local inference practical on much smaller devices.

Two low-bit operating points

Variant	Weight representation	Deployed language-model footprint	Focus
1-bit Bonsai 27B	Binary {−1, +1}, 1.125 effective bits per weight	~3.9 GB	Smallest footprint
Ternary Bonsai 27B	Ternary {−1, 0, +1}, 1.71 effective bits per weight	~7.2 GB	Higher quality retention

PrismML applies the low-bit representation end to end across the language network, including embeddings, attention, MLPs, and the LM head. The headline footprints describe the language-model weights; actual runtime memory also includes the KV cache, activations, runtime buffers, and the optional vision tower, and grows with context length.

Capabilities

• 262K-token context: supports long documents and large working sets.
• Vision: accepts image input through a separately packaged 4-bit vision tower.
• Reasoning: reasons by default, with a configurable thinking budget.
• Tool use: supports structured function calling through the standard chat API.

Benchmark results

PrismML reports the following category averages from 15 benchmarks evaluated in thinking mode:

Category	Qwen3.6 27B FP16	Ternary Bonsai 27B	1-bit Bonsai 27B
Math	95.3	93.4	91.7
Coding	88.7	86.0	81.9
Agentic and tool calling	80.0	74.0	66.0
Instruction following	78.4	71.8	65.8
Knowledge / STEM	83.1	77.0	73.4
Vision	72.6	65.2	59.6
Overall (15 benchmarks)	85.0	80.5	76.1

Benchmark results and chart are reported by PrismML. See the Bonsai 27B announcement and whitepaper for the full per-benchmark results and methodology.

License

Bonsai 27B is available under the Apache 2.0 License.