PrismML unveils Bonsai 27B, a compressed multimodal 27B-class AI model designed to run locally on phones

PrismML announced Bonsai 27B, based on Qwen3.6 27B, with ternary and 1-bit variants that shrink the model to 5.9GB and 3.9GB respectively. The company says the model supports multimodal tasks, long 262K-token context, tool calling, reasoning, and agentic computer-use workflows on local devices including laptops and the iPhone 17 Pro.

Bonsai 27B highlights how aggressive low-bit quantization is pushing frontier-like AI capabilities toward consumer hardware instead of relying solely on cloud inference. If performance claims hold up, this could accelerate private, offline AI assistants and pressure device makers to optimize chips and memory around local model execution.