PrismML Releases Bonsai 27B, a 1-Bit Model That Runs Locally on an iPhone

PrismML released Bonsai 27B on July 14, an open-weights model built on Qwen3.6 27B that the company says is the first 27B-class model small enough to run locally on a phone. The trick is quantization-aware training baked in from the start, rather than compressing a model after the fact.

• 1-bit variant: 3.9GB (1.125 effective bits/weight) vs 54GB at standard 16-bit precision
• Ternary variant: 5.9GB (1.71 effective bits/weight), aimed at laptops
• 11 tokens/second on iPhone 17 Pro; up to 163 tokens/second on an RTX 5090
• 262K-token context window, multimodal with a compact 4-bit vision tower
• Released under Apache 2.0, plus a free limited-time developer preview API

They're really evaluating our technology right now, characterizing the discussions as very early - PrismML CEO Babak Hassibi, on Apple's interest in the model

The release lands as open-weights models keep pushing into territory once reserved for closed, hosted APIs. See how Bonsai 27B and other open-weights entrants compare head-to-head in our coding arena.