PrismML, a U.S. AI company, has added three new models to its Bonsai 2 family: 2B, 4B, and 9B, all built on Alibaba’s Qwen3.5. The team said it constrained the main weights of the language models to just three values, -1, 0, and +1, reducing the model files to 0.43GB, 0.95GB, and 2GB, or about one-ninth the size of the original 16-bit weights. The company said the new models still support reasoning, image understanding, and tool use, with a focus on local agents running on phones and personal computers. According to PrismML, the 9B model can generate up to 34 tokens per second on an iPhone 18 Pro Max, while the 2B version reaches 120 tokens per second. In the company’s official tests, the 9B model came close to the original same-sized Qwen3.5 in coding and instruction-following tasks, but showed a wider gap on more complex workloads. PrismML has released the weights for all three models under the Apache 2.0 license.
PrismML, a U.S. AI company, has added three models to its Bonsai 2 family: 2B, 4B, and 9B. All three are based on Alibaba’s Qwen3.5. The team said it limited the main language model weights to three values, -1, 0, and +1, cutting file sizes to 0.43GB, 0.95GB, and 2GB, roughly one-ninth the size of the original 16-bit weights.
The new models still support reasoning, image understanding, and tool calling, and are aimed mainly at local agents running on devices such as smartphones and personal computers. PrismML said the 9B model can generate as many as 34 tokens per second on an iPhone 18 Pro Max, while the 2B version reaches 120 tokens per second.
Official benchmark results
In PrismML’s official tests, the 9B model came close to the original same-sized Qwen3.5 in coding and instruction-following ability, though the gap was more visible on complex tasks. On Tau2Bench, which tests multi-turn dialogue and tool use, Bonsai 2 9B scored 73.02, compared with 85.25 for the original model. On complex document parsing, it scored 73.8, versus 88.73 for the original.
The weights for all three models have been released under the Apache 2.0 license.
This article was originally published by Bit.Fan. For more cryptocurrency news and market insights, visit www.bit.fan. Disclaimer:
The market information, project data, and third-party content displayed on this platform are for industry information sharing only and do not constitute any form of investment advice or return commitment.
Cryptocurrency trading carries high risks. Users should fully assess their risk tolerance and make independent decisions. All profits, losses, and legal responsibilities are borne by the users themselves.
