PrismML Packs Its 9B Bonsai 2 Model Into 2GB for Local AI

PrismML released three Bonsai 2 open-weight models for local use, including a 9B version with 2GB of weights, trading smaller files for weaker scores on complex benchmarks.

The 2B, 4B and 9B models are based on Alibaba’s Qwen3.5. Their main language-model weights are restricted to three values — −1, 0 and +1 — bringing the files to 0.43GB, 0.95GB and 2GB, respectively, about one-ninth the size of the original 16-bit weights.

All three models support reasoning, image understanding and tool use, and are designed for local agents on phones and personal computers. On an iPhone 18 Pro Max, the 9B model generates up to 34 tokens per second, while the 2B model reaches up to 120. The weights are available under the Apache 2.0 license.

Bonsai 2 9B scored 73.02 in Tau2Bench, which tests multi-turn conversations and tool use, compared with 85.25 for same-scale original Qwen3.5. On complex document parsing, Bonsai 2 9B scored 73.8, while Qwen3.5 scored 88.73.