Non-uniform Quantization
A quantization method that dynamically allocates bit widths according to the actual distribution of weights. Important weights retain higher precision while long-tail weights are compressed to very low bit widths, thereby keeping error small at extremely low bit widths.
Timeline (last 90 days)
A European team released an extreme quantization version of DeepSeek, compressing the 763-billion-parameter model to an average of 1.75 bits with near-zero loss on mainstream benchmarks.
Non-uniform quantization dynamically allocates bits according to weight distribution, keeping more precision for high-frequency important weights and compressing long-tail weights to extremely low bits, thereby maintaining small error even at very low average bit width.
All Facts (2)
A European team released an extreme quantization version of DeepSeek, compressing the 763-billion-parameter model to an average of 1.75 bits with near-zero loss on mainstream benchmarks.
50%UnverifiedNon-uniform quantization dynamically allocates bits according to weight distribution, keeping more precision for high-frequency important weights and compressing long-tail weights to extremely low bits, thereby maintaining small error even at very low average bit width.
50%