PrismML's Ternary Bonsai 2 27B Brings Extreme AI Compression to the Edge

PrismML, a Caltech-founded AI lab, has released Bonsai 2 27B, a 27.8-billion-parameter multimodal model compressed to just 5.9 GB that retains 98.2% of full-precision performance while enabling reasoning, coding, vision, and agentic tasks to run directly on consumer devices. Available freely under the Apache 2.0 license since September 17, 2026, the ternary-quantized model represents a watershed moment in on-device AI deployment—running the Qwen3.8 27B architecture with more than 9× memory reduction compared to standard implementations.