Bonsai 27B
A 27B-class compressed multimodal model for developers who want large-model capability on phones or constrained local devices, helping them build offline reasoning, tool-use, and agent workflow prototypes.
Tool overview
Based on the current evidence, Bonsai 27B looks worth tracking as a serious on-device large-model compression effort, but not yet something we can confidently treat as a broadly proven product. The adoption judgment is cautiously positive: PrismML shared materials for phone execution, a WebGPU demo, an end-to-end agent workflow, and benchmark claims across 15 tasks, while also stating the model is based on Qwen 3.6 27B and released in binary and ternary variants under Apache 2.0. Still, almost all evidence here comes from PrismML posts or repost-style social spread, so attention is strong while usability proof is still mostly first-party.