AI Paper Atlas

Research map / Architectures and efficiency / Quantization

MixFP4: Enhancing NVFP4 with Adaptive FP4/INT4 Block Representations

Jiaxiang Zou, Yonghao Chen, Ruilong WU, Xinyu Chen

ICML 2026 · Architectures and efficiency › Quantization · approach: Ultra-low-bit LLM quantization · 0 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
use MixFP4 per-block E2M1/E1M2 selection instead of NVFP4NVFP4 RTN · WikiText, Llama-3.2-1Bperplexity: 13.90 -> 13.48Improved
use MixFP4 instead of NVFP4NVFP4 RTN · WikiText, Llama-3.1-8Bperplexity: 8.26 -> 8.06Improved
use MixFP4 instead of 4/64/6 RTN · WikiText, Qwen3-8Bperplexity: 12.56 -> 12.39Improved
use MixFP4 instead of NVFP4 on Qwen3-4B with RHTNVFP4 +RHT · WikiText, Qwen3-4Bperplexity: 19.22 -> 18.60Improved
use MixFP4 instead of NVFP4 on Qwen3-4B without RHTNVFP4 RTN · WikiText, Qwen3-4Bperplexity: 17.34 -> 17.42No clear effect
add random Hadamard transform to NVFP4 (ablation)NVFP4 RTN · WikiText, Llama-3.1-8Bperplexity: 8.26 -> 8.55Hurt
add random Hadamard transform to MixFP4 (ablation)MixFP4 RTN · WikiText, Llama-3.1-8Bperplexity: 8.06 -> 8.10Hurt
use MixFP4 with SpinQuant instead of NVFP4SpinQuant · Qwen3-8B avg accuracyaverage accuracy: 67.20 -> 69.78Improved
use MixFP4 with GPTQ instead of 4/6GPTQ · Llama-3.1-8B wikiperplexity: 8.16 -> 8.05Improved
add E1M2 format option to FP4-E2 at block size 16 (ablation)NVFP4 · Llama-3.1-8B WikiTextperplexity: 8.26 -> 8.06Improved
add E3M0 format option to FP4-E2 at block size 16 (ablation)NVFP4 · Llama-3.1-8B WikiTextperplexity: 8.26 -> 8.21Improved
add E1 and E3 options together at block size 16 (ablation)NVFP4 · Llama-3.1-8B WikiTextperplexity: 8.26 -> 8.01Improved

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy