An update to a quantization compiler replaced k-means with e-means in codebook generation. This reduced the Qwen3.5-4B model's size from 8.41 GB to 1.23 GB while retaining 91.9% of its original performance, comparable to bonsai 2.
E-means Quantization Compiler Update Shrinks Qwen3.5-4B, Retains PerformanceAI News
An update to a quantization compiler replaced k-means with e-means in codebook generation. This reduced the Qwen3.5-4B model's size from 8.41 GB to 1.23 GB while retaining 91.9% of its original performance, comparable to bonsai 2.