PrismML Ships Bonsai 27B: 1-bit and ternary Qwen3.6-27B Bring On-Device AI
PrismML released Bonsai 27B, a compressed, non-pretrained version of Qwen3.6-27B in two low-bit formats: ternary weights at 1.71 bits/weight (~5.9GB) and 1-bit weights at 1.125 bits/weight (~3.9GB). Both variants are multimodal and maintain the same architecture, with memory and KV cache constraints shaping device viability. In benchmarking across 15 tests on H100, ternary Bonsai preserves about 94.6% and 1-bit about 89.5% of FP16 accuracy, and with additional 4-bit KV cache it can fit on devices; PrismML notes on-device usage for laptops and phones, speedups via DSpark, and Apache 2.0 licensing with llama.cpp/MLX support.
