Bonsai 2 27B

Bonsai 2 27B, released September 17, 2026 by Pasadena startup PrismML, compresses Alibaba's Qwen3.8 27B into ternary weights — every weight restricted to -1, 0, or +1, about 1.72 bits per weight with grouping. The result is a 27.8B-parameter multimodal model in 5.9GB that PrismML says retains 98.2% of the base model's aggregate benchmark performance across a 20-benchmark suite covering reasoning, math, coding, instruction-following, vision, and agentic tool use. It ships Apache 2.0 weights, runs at up to 143 tokens/second on an RTX 5090, and carries a 262K-token context window. Founder and CEO Babak Hassibi, a Caltech professor, leads the company, which is backed by Khosla Ventures.

The honest picture: 98.2% is PrismML's own number on PrismML's own suite — independent third-party benchmarking is still pending, and the company keeps a thin public profile. Ternary quantization is genuinely novel, but real-world agentic workloads may expose gaps that aggregate benchmarks smooth over. Still, the named source model, disclosed methodology, and specific retention figure make this a more disciplined compression claim than most.