gg2
Bonsai 2 27B near-lossless compression model
The gist
PrismML released Ternary Bonsai 2 27B, a 5.9GB model keeping 98.2% of Qwen3.8 27B's benchmark score.
The tradeoff is uneven: vision and tool-calling lose the most.
How it unfolded
- Jul 15, 2026GIGAZINE reported the previous Bonsai 27B release, a smaller version of Qwen3.6-27B using only 3.9GB of memory and compatible with iPhones.
- Sep 17, 2026PrismML released Ternary Bonsai 2 27B, based on Qwen3.8 27B, with ternary {-1, 0, +1} weights, FP16 group-wise scaling, 1.76 effective bits per weight, a 5.9GB footprint, a 262K-token context window, multimodal text-and-image input, and an Apache 2.0 license.
- Sep 17, 2026PrismML reported the model scores 83.9 overall versus 85.4 for Qwen3.8 27B and 83.6 for Qwen3.6 27B, and said it is more than 9x smaller than its full-precision counterpart while retaining 98.2% of aggregate benchmark performance.