PrismML expands Bonsai 2 with three compact AI models for local agents

American AI company PrismML has expanded the Bonsai 2 family, introducing three new models with 2, 4 and 9 billion parameters. All of them are based on Qwen 3.5 from Alibaba.

The developers restricted the main weights of the language models to three values: −1, 0 and +1. This allowed file sizes to be reduced to 0.43 GB, 0.95 GB and 2 GB, respectively. That is roughly nine times smaller than the original models with 16-bit weights.

The new models still support reasoning, image understanding and tool calling. They are primarily intended for local AI agents running on smartphones, personal computers and other devices.

According to PrismML, on the iPhone 18 Pro Max the 9B model generates up to 34 tokens per second, while the 2B version reaches 120 tokens per second.

In official tests, Bonsai 2 9B is nearly on par with the original Qwen 3.5 model of the same size in coding and instruction following, but the gap becomes more noticeable on complex tasks.

In Tau2Bench, which evaluates the ability to conduct multi-step dialogs and work with tools, Bonsai 2 9B scored 73.02 points versus 85.25 for the original Qwen 3.5. In a test on analyzing complex documents, the results were 73.8 and 88.73 points, respectively.

Weights for all three models have already been published under the Apache 2.0 license.

At the same time, it is worth noting that 2 GB is the size of only the 9B language model in the most compact format. For image processing, an additional visual module of about 0.62 GB is required.

Model collection on Hugging Face