AI · 1h ago
Bonsai 27B shrinks 54GB AI model to 4GB for phones
Prism ML compressed Qwen 3.6 from 54GB to under 4GB using 1-bit quantization, fitting on a phone. In tests, the compressed model lost factual accuracy (0% correct on specific dates) but matched the original on coding tasks. The ternary version (7.2GB) also showed severe factual degradation.
Meridian48 take
The trade-off is stark: extreme compression preserves coding ability but destroys factual recall, making this a niche tool for developers, not a general assistant.
Read the full reporting
Testei o Bonsai 27B, o modelo que cabe no seu telefone. Vale a pena? →
DEV Community
model-compressionon-device-ai