27.3B language parameters + ~0.46B vision tower
262K tokens
Apache 2.0
Official open-weight Bonsai 27B release family from Prism ML, positioned around extreme intelligence density and on-device deployment rather than conventional full-precision scaling. The public Prism ML materials describe Bonsai 27B as the newest and largest generation in the Bonsai line and the family’s first 27B-class vision-language models, derived from Qwen3.6-27B with the architecture kept unchanged: a hybrid-attention causal model with roughly 75% linear and 25% full attention, SwiGLU MLPs, RoPE, RMSNorm, a 64-block backbone, and an optional ~0.46B vision tower. The release ships multiple public operating points, including native 1-bit Bonsai, ternary Bonsai, AWQ 4-bit, MLX, GGUF, and unpacked safetensors variants. Prism ML says the 1-bit Bonsai 27B model uses binary g128 weights with FP16 group-wise scaling, compressing the deployed language model to about 3.9 GB, while the ternary quality-oriented variant deploys at about 7.2 GB and retains more of the full-precision model’s behavior. Public release materials emphasize 262K-token context, near-lossless 4-bit KV-cache quantization, native OpenAI-style tool calling, MCP support in the demo stack, multimodal image understanding for the 27B models, per-request reasoning effort control, and optional DSpark speculative decoding drafters. The release is framed as the first 27B-class model family able to run on a phone in its native 1-bit format, with published claims of roughly 11 tok/s on an iPhone 17 Pro Max and about 44 tok/s on an M5 Pro laptop, while the ternary family is positioned as the higher-quality default operating point. The public whitepaper and model cards present the family as an efficiency-first, long-context, agent-friendly alternative for local and edge deployment through MLX, llama.cpp, Transformers, vLLM, and SGLang-compatible workflows.