IT Home On October 16, a news report stated that AI company PrismML posted a blog post yesterday (October 15) announcing the release of the Ternary Bonsai 2 series of models. These models come in three sizes: 2B, 4B, and 9B. The goal is to enable local operation of AI agents on smartphones, PCs, and wearable devices, allowing them to perform reasoning, image recognition, and tool invocation.
Ternary Bonsai 2 series models are built based on Alibaba’s Qwen 3.5 base, using a “ternary” weighting method, where model parameters are encoded with only three values: -1, 0, and +1. According to the official data, the storage requirements for the 2B, 4B, and 9B models are approximately 0.43GB, 0.95GB, and 2GB respectively.
In terms of capabilities, the Bonsai 2 series emphasizes the practicality of local agents: following instructions, completing daily tasks, invoking tools on mobile devices, and avoiding redundant output. The official tests evaluated six capabilities—smart tool invocation, mathematics, instruction following, programming, knowledge and reasoning, and visual understanding—on 19 benchmarks. The 9B model reduced the number of tokens required for standard instruction execution and daily mathematical tasks by approximately 40%.
In terms of performance, PrismML states that on hardware such as the RTX 4090 and M5 Pro, the new model can achieve up to 5 times higher throughput and 3 times higher energy efficiency compared to the full-precision model. On the iPhone 18 Pro Max, the 9B model can reach up to 34 word elements per second, and the 2B model can reach up to 120 word elements per second. IT Home has provided the relevant screenshots as follows:
This series can form a complete product line ranging from lightweight tasks to complex workflows, in conjunction with the existing Bonsai 2 27B. Developers can run it on NVIDIA GPUs using llama.cpp, or deploy it on Mac, iPhone, and iPad using MLX; the model weights are available under the Apache 2.0 license.