GoKawiilBonsai 2 27B compresses a 27B model to 5.9GB with 98.2% performance retained
A new ternary-quantized model, Ternary Bonsai 2 27B, has been released, built on Qwen3.8 27B and using {-1,0,+1} weights with FP16 group scaling to shrink the model to about 1.76 effective bits per weight and a 5.9GB footprint. Despite being over 9x smaller than its full-precision counterpart, it retains 98.2% of aggregate benchmark performance across reasoning, coding, vision and agentic tasks, and supports a 262K-token context window under an Apache 2.0 license.
GoKawiil's interpretation of the reporting above, not reported fact.
This level of compression turns a large multimodal model into something that can realistically run on local devices with limited memory and power, rather than requiring cloud infrastructure. Near-lossless performance at 9x smaller size suggests ternary quantization is maturing into a practical technique for on-device AI deployment, not just a research curiosity.
- Ternary Bonsai 2 27B shrinks a 27B-parameter model to 5.9GB while keeping 98.2% of its benchmark performance.
- It is based on Qwen3.8 27B and adds stronger reasoning, coding, vision, and agentic capabilities over the prior Bonsai release.
- The model supports a 262K-token context window, multimodal image-text input, and is released under Apache 2.0.
Source: prismml.com, 2026-09-17
Published there as: “Bonsai 2 27B: Near-Lossless Compression in a 9x Smaller Footprint”
Read the original report → The summary and analysis above are GoKawiil's own, written from reporting by the source above. Facts and quotes belong to the original publisher.