AI BriefingWarpQuant: 3-bit Quantization Based on Hadamard Rotation
·2026.08.16 10:58
Key point
WarpQuant, a 3.62 bpw quantization method combining Hadamard rotation and Output-Fisher recovery to enhance Qwen 27B performance, has been released.
Details
WarpQuant is a new post-training quantization (PTQ) methodology that separates the compression domain and the recovery domain.
Key Technical Features:
- Hadamard Rotation: Applies a deterministic sign Hadamard rotation ($R=HD$) to projected weights.
- 3-bit Quantization: Quantizes rotated weights using a 3-bit Gaussian Lloyd-Max codebook and block-GPTQ error feedback.
- Output-Fisher Recovery: Recovers the rank of the most sensitive columns using Output-Fisher sensitivity calculated from next-token NLL gradients.
Qwen3.8-27B Performance Results:
- Achieved a low bit rate of 3.6165 bpw while recording superior PPL (7.4737) and ARC (56.86%) performance compared to the existing IQ3_S method.
- Effectively maintained benchmark performance while drastically reducing capacity compared to BF16.
Other Models and Deployment:
- Demonstrated performance improvements over existing quantization methods on Qwen3.5-4B and Llama 3 8B models.
- Currently provides checkpoints in safetensors format via Hugging Face, with GGUF format support for Ollama planned for future release.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
Stanford and NVIDIA Release 'Contrastive Language Model' CLM with 9x Faster Inference