Prism ML
An AI research company focused on extreme model compression techniques. Prism ML developed the Bonsai series of ternary-quantized large language models, applying aggressive weight quantization methods to reduce model size while maintaining usable performance.
Timeline (last 90 days)
The Bonsai 2 model released by Prism ML claims to retain 98% of the performance of a 27-billion-parameter model within less than 8 GB of VRAM.
Before converting to ternary values, Prism ML applies a rotation matrix (orthogonal transformation) to smooth and disperse weight outliers so that the shared scale is more appropriate, an idea drawn from research such as BitNet and QuIP#.
An ordinary 2-bit quantized version is approximately 7.3 GB and can only retain approximately 88% of the benchmark score.
Running Bonsai 2 on a MacBook is recommended using Prism ML's llama.cpp fork, as it caches conversations between steps, whereas the MLX version does not cache and is noticeably slower.
Bonsai 2 is a ternary-quantized large model released by Prism ML, retrained from scratch on a Qwen3-27B base, with an overall file size of approximately 6 GB.
Prism ML claims that Bonsai 2 retains 98% of the original Qwen model's performance after aggressive compression.
All Facts (5)
The Bonsai 2 model released by Prism ML claims to retain 98% of the performance of a 27-billion-parameter model within less than 8 GB of VRAM.
50%UnverifiedBefore converting to ternary values, Prism ML applies a rotation matrix (orthogonal transformation) to smooth and disperse weight outliers so that the shared scale is more appropriate, an idea drawn from research such as BitNet and QuIP#.
50%UnverifiedAn ordinary 2-bit quantized version is approximately 7.3 GB and can only retain approximately 88% of the benchmark score.
50%UnverifiedBonsai 2 is a ternary-quantized large model released by Prism ML, retrained from scratch on a Qwen3-27B base, with an overall file size of approximately 6 GB.
50%UnverifiedPrism ML claims that Bonsai 2 retains 98% of the original Qwen model's performance after aggressive compression.
50%