PrismML launches 1.58-bit model Ternary Bonsai, with parameters reduced by 9 times, surpassing peers in intelligence

ME News, April 17 (UTC+8). According to Dongcha Beating monitoring, PrismML has released the Ternary Bonsai series of language models. Using 1.58-bit (ternary weights) technology, the models reduce VRAM usage to one-ninth of that of a 16-bit model while maintaining high performance. The series includes three parameter sizes: 8B, 4B, and 1.7B. They are now open-sourced on Hugging Face and support native operation on Apple devices.

The so-called 1.58-bit model means restricting neural network weights to three values: {-1, 0, +1}. Compared with the previously pursued ultra-compressed 1-bit models (weights only {-1, +1}), introducing the “0” value can effectively remove redundant connections, allowing the model to preserve complex reasoning capabilities at an extremely small size. The released Ternary Bonsai 8B weight file is only 1.75 GB, with a benchmark average score of 75.5. This is not only 5 points higher than the company’s own 1-bit version, but also significantly leads in “intelligent density” (performance contributed per GB of VRAM) over similar dense models such as Qwen3.

Energy efficiency and operating speed are another core advantages of this series. On the iPhone 17 Pro Max, the 8B version can reach a speed of 27 tok/s, with an energy-efficiency improvement of about 3 to 4 times. For developers who need to deploy high-performance AI on edge devices such as phones and laptops, this means achieving intelligent performance close to that of full-precision models at a very small memory cost.

Currently, the Ternary Bonsai models are natively supported on Apple devices via the MLX framework. The model weights are distributed under the Apache 2.0 license.
(Source: BlockBeats)