PrismML has released Bonsai 2 in 2B, 4B, and 9B sizes, with model files as small as 430MB. The models are based on Qwen3.5 and use ternary quantization to make local AI more practical on devices with limited memory. The model sizes: Bonsai 2 2B: 430MB Bonsai 2 4B: 950MB Bonsai 2 9B: 2GB All three are released under Apache 2.0, allowing commercial and personal use under the license terms. But the interesting part isn’t just how small these models are. PrismML is also targeting the capabilities that make an AI assistant useful beyond ordinary text generation. What has improved? Bonsai 2 is designed to improve reasoning, visual understanding, instruction following, and tool calling. The team has also worked on reducing repetitive responses, unnecessary verbosity, and output loops. Those issues can make smaller models frustrating to use, particularly in agentic workflows where a model needs to follow instructions and interact with external tools. PrismML reports that Bonsai 2 9B uses roughly 40% fewer tokens on standard instruction-following and everyday math tasks. That could reduce generation time and inference costs. However, fewer tokens are only an advantage if the model still produces correct, complete answers. Actual speed also depends on the hardware and inference runtime. How does ternary quantization work? Most conventional models represent their weights using several bits per parameter. Ternary quantization reduces the possible weight values to three choices, typically -1, 0, and +1, with scaling factors used to preserve useful weight magnitudes. This can dramatically reduce storage requirements compared with conventional weight formats. The result is a 9B model whose weight file is approximately 2GB, opening up possibilities for running more capable models on memory-constrained hardware. One important distinction: the weight file is not the full memory requirement. Inference also needs memory for the KV cache, intermediate computations, and runtime overhead. The actual hardware requirements will depend on the implementation and context length. Why this matters for local AI agents PrismML is positioning Bonsai 2 for applications on personal computers, smartphones, and wearables. Imagine an assistant that follows instructions, interprets visual information, and calls tools locally rather than sending every interaction to a remote server. That could help reduce dependence on cloud inference, improve privacy for suitable workloads, and make AI features available on devices with limited resources. The Bonsai family now spans four sizes, from 2B to 27B parameters, giving developers more options when balancing model capability against hardware constraints. Bonsai 2 also builds on the earlier 1.7B, 4B, and 8B models. The next question is how reliably these compact models perform on real tasks. Tool-call accuracy, reasoning quality, visual understanding, and resistance to repetitive output will matter more than file size alone. Availability: Bonsai 2 2B, 4B, and 9B are released under Apache 2.0.