News
Ternary Bonsai 2 series AI model debuts: local throughput on Apple iPhone 18 Pro Max reaches up to 120 words per second
2 min read
Source: ithome.com
IT House reported on October 16th that AI company PrismML published a blog post yesterday (October 15th), announcing the launch of the Ternary Bonsai 2 series of models, which are available in three sizes: 2B, 4B and 9B. The goal is to run AI agents locally on mobile phones, PCs and wearable devices to perform reasoning, image recognition, tool calling and other capabilities. The Ternary Bonsai 2 series models are built on Alibaba's Qwen 3.5 base and adopt the "ternary" weight representation method, that is, only three numerical encoding model parameters -1, 0, and +1 are used. According to officials, the storage footprint of the 2B, 4B, and 9B models is 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, calling tools on mobile devices, avoiding repeated output, etc. Officials tested six categories of abilities: intelligent tool invocation, mathematics, instruction following, programming, knowledge and reasoning, and visual understanding on 19 benchmarks; among them, the 9B model reduced the number of tokens required by approximately 40% in standard instruction execution and daily mathematics tasks. In terms of performance, PrismML said that on hardware such as RTX 4090 and M5 Pro, the new model can achieve up to 5 times the throughput and 3 times the energy efficiency compared to the corresponding full-precision model; on the iPhone 18 Pro Max, the 9B model can reach up to 34 words per second, and the 2B model can reach up to 120 words per second. IT Home has attached relevant screenshots as follows: This series can be combined with the existing Bonsai 2 27B to form a complete product line from light tasks to complex workflows. Developers can run on NVIDIA GPUs via Llama.cpp and deploy on Mac, iPhone and iPad via MLX; model weights are open under the Apache 2.0 license.