PrismML's one-bit Bonsai 27B runs on an iPhone

On July 14, 2026, PrismML released one-bit and ternary Bonsai 27B variants based on Qwen3.6-27B under Apache 2.0. The 3.9 GB one-bit build uses 1.125 bits per weight and retained 89.5% of FP16 performance across 15 benchmarks. It supports a 262,000-token context and multimodal input through a vision tower. PrismML reported up to 87 tokens per second on M5 Max and 163 on RTX 5090.

Visit original source