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

ME News message. On April 17 (UTC+8), according to Dongcha Beating monitoring, PrismML released the Ternary Bonsai series of language models. Using 1.58-bit (ternary weights) technology, the models reduce VRAM usage to one-ninth of a 16-bit model while maintaining high performance.

The series includes three parameter sizes: 8B, 4B, and 1.7B. They have now been open-sourced on Hugging Face and support native execution on Apple devices.

The so-called 1.58-bit model means restricting neural network weights to three values: {-1, 0, +1}. Compared with the previous 1-bit models that pursued ultra-extreme compression (weights only {-1, +1}), introducing the “0” value can effectively remove redundant connections, allowing the model to retain complex reasoning capabilities in an extremely small footprint.

The released Ternary Bonsai 8B weight file is only 1.75 GB, and its average benchmark score reaches 75.5. This is not only 5 points higher than its own 1-bit version, but also significantly outperforms similar dense models such as Qwen3 in “intelligence density” (performance contributed per GB of VRAM).

Energy efficiency and runtime speed are another core advantage of this series. On the iPhone 17 Pro Max, the 8B version can achieve speeds of up to 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 being able to obtain intelligent performance close to full-precision models at the cost of very little memory.

Currently, the Ternary Bonsai models are already supported natively on Apple devices through the MLX framework. Model weights are distributed under the Apache 2.0 license.

(Source: BlockBeats)