bonsai-27b

Bonsai 27B

Bonsai 27B is PrismML's family of binary and ternary Qwen3.6 27B models, designed to retain strong reasoning, coding, and agentic capabilities at exceptionally small footprints.

Highlights

• 27B reasoning in a phone-class footprint: binary weights compress the language model to about 3.9 GB, bringing 27B-class local inference to high-end phones.
• Quality that survives extreme compression: across 15 thinking-mode benchmarks, the binary model retains 89.5% of FP16's average score at 1.125 effective bits per weight; the ternary model retains 94.6% at 1.71 bits.
• Ready for demanding local workflows: supports vision input, tool use, hybrid thinking, and a 262K-token context window.