I used to think that running large models on mobile phones was just a gimmick. Now, with 8B achieving a score of 75.5 and 27 tokens per second, I need to reevaluate.

PrismML releases the Ternary Bonsai series, using 1.58-bit weights {-1, 0, +1}, with GPU memory only one-ninth of a 16-bit model. The 8B/4B/1.7B sizes are open-sourced on Hugging Face and natively run on Apple devices. The 8B weights are approximately 1.75 GB, with a benchmark score of 75.5, leading among peers. On the iPhone 17 Pro Max, 8B runs at 27 tokens/sec, with a 3–4 times improvement in energy efficiency. The weights are distributed under Apache 2.0 and run natively on Apple devices via the MLX framework.