Bonsai 27B model compresses Qwen 3.6 to phone-size, excelling at code tasks

A new model called Bonsai 27B, developed by Prism ML, a spinout from Caltech, has been created by aggressively compressing the Qwen 3.6 model. This compression allows the model to fit on a phone, reducing its size from approximately 54GB to under 4GB. While the Bonsai 1-bit and ternary versions struggle with factual recall, they perform comparably to the original model on coding tasks and even show promise in creative code generation. AI

IMPACT Enables on-device AI applications by drastically reducing model size, though with trade-offs in factual knowledge.

RANK_REASON The item discusses a compressed model that can run on a phone, which is a practical application or tool, rather than a frontier release or significant research.

Read on dev.to — LLM tag →

• Bonsai 27B
• California Institute of Technology
• Prism ML
• Qwen 3.6

• model release
• product

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →