PrismML Releases Bonsai 27B — First 27B-Class Model to Run on a Phone

What happened

PrismML (backed by Khosla Ventures and Google, founded by Caltech researchers) released Bonsai 27B on July 14, 2026 — the first 27B-parameter model capable of running directly on a mobile phone. Apple is reportedly actively evaluating the technology.

Context and impact

Until now, 27B-parameter models were exclusive to server environments. Bonsai 27B brings frontier-class capabilities (coding, reasoning, agentic workflows, vision) directly on-device without requiring cloud connectivity. Apache 2.0 licensing enables commercial deployment at zero cost.

Details

• 1-bit variant: 3.9 GB, 11 tokens/s on iPhone 17 Pro, retains 90% of full-precision performance
• 1.58-bit ternary variant: 5.9 GB, retains 95% of full-precision performance
• Context window: 262,000 tokens
• Capabilities: text, code, vision, structured tool calling, reasoning
• License: Apache 2.0 — free for commercial and non-commercial use
• Base: built on Qwen3.6 27B