PrismML Releases Bonsai 27B — Extreme Compression for Local AI
PrismML has released Bonsai 27B, a heavily compressed 27B-parameter model designed to run where models of its class were previously impractical: laptops, local agents, and even phones. Built on Qwen3.6 27B, it comes in two flavors — Ternary Bonsai 27B (5.9 GB, ~1.71 effective bits/weight) for desktops and laptops, and 1-bit Bonsai 27B (3.9 GB, ~1.125 effective bits/weight) small enough to fit the memory budget of an iPhone 17 Pro.
Despite the aggressive compression, the model retains vision, tool calling, agentic loops, structured outputs, and 256k+ context, with support for speculative decoding, MLX (Apple), CUDA, and llama.cpp. Benchmarks show ~95% of full-precision quality for the ternary version and ~90% for the 1-bit version across reasoning, math, coding, instruction following, tool calling, and vision tasks. Released under Apache 2.0 with models on Hugging Face, a WebGPU demo, API via Together, and a whitepaper.