PrismML Adds 2B, 4B and 9B Models to Bonsai 2

PrismML announced in a press release three Bonsai 2 ternary models with 2 billion, 4 billion and 9 billion parameters. Based on Qwen3.5, the models support reasoning, vision and tool calling on personal computers, phones, smart glasses and other wearables.

The models use ternary weights and have respective memory footprints of 0.43GB, 0.95GB and 2GB. PrismML says they matched or exceeded the corresponding full precision Qwen3.5 models on coding, math and instruction following across its benchmark suite, while retaining more than 90% of their scores in vision, tool use, knowledge and reasoning tasks.

PrismML reports throughput of up to 276 tokens per second for Bonsai 2 2B and 124 tokens per second for Bonsai 2 9B on desktop hardware. On an iPhone 18 Pro Max, the models reached up to 120 and 34 tokens per second, respectively.

The three models are available for free under the Apache 2.0 license. PrismML trained them using Google v5 TPUs and optimized them for consumer CPUs and edge GPUs.