Research map / Architectures and efficiency / Quantization
MicroMix: Efficient Mixed-Precision Quantization with Microscaling Formats for Large Language Models
Wenyuan Liu, Haoqian Meng, Yilun Luo, Peng Zhang, Xindian Ma
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 changed | Where | Measured | Effect |
|---|---|---|---|
| use MXFP8 E4M3 and MXFP6 E3M2 instead of E5M2 and E3M2 (ablation) | MicroMix · WikiText2, Llama3.1-8B | perplexity: 6.84 -> 6.72 | Improved |
| use MXFP6 E3M2 instead of E2M3 with MXFP8 E4M3 (ablation) | MicroMix · Llama3.1-8B zero-shot and WikiText2 | Winogrande: 71.51 -> 72.14 | Mixed |
| use MXFP6 E2M3 instead of E3M2 with MXFP8 E5M2 (ablation) | MicroMix · WikiText2, Llama3.1-8B | perplexity: 6.84 -> 6.81 | No clear effect |
| calibrate on Pile instead of WikiText2 (ablation) | MicroMix · WikiText2, Qwen2.5-14B | perplexity: 5.87 -> 5.92 | No clear effect |
| calibrate on C4 instead of WikiText2 (ablation) | MicroMix · WikiText2, Qwen2.5-14B | perplexity: 5.87 -> 5.92 | No clear effect |
| apply Atom or QUIK fixed-channel mixed precision to MXFP formats instead of MicroMix thresholds (ablation) | Atom on MXFP · WikiText2, Llama3.1-8B | perplexity: 6.72 -> 8.02 | Hurt |
| vary average bit-width from 4.48 to 6.01 (ablation) | MicroMix · Qwen2.5-32B zero-shot avg | accuracy: 74.89 -> 75.20 | No clear effect |
| replace FP16 with MicroMix | Mixtral-8x7B-Instruct · zero-shot avg | accuracy: 78.58 -> 78.20 | No clear effect |
| replace FP16 with MicroMix | Mixtral-8x7B-Instruct · execution time | time: 5min 18s -> 2min 03s | Improved |
| quantize KV cache to INT4 on top of MicroMix | MicroMix · Llama3.1-8B WikiText2 | perplexity: 6.72 -> 6.97 | Hurt |
| use MicroMix instead of 6-bit INT quantization | INT6 · Llama3.1-8B MMLU | accuracy: 58.67 -> 62.65 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- FlatQuant — FlatQuant: Flatness Matters for LLM Quantization (ICML 2025) · better on 6 of 6 shared results
- QuaRot — QuaRot: Outlier-Free 4-Bit Inference in Rotated LLMs (NeurIPS 2024) · better on 3 of 3 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Progressive Mixed-Precision Decoding for Efficient LLM Inference (ICLR 2025 · similarity 0.89)
- SqueezeLLM: Dense-and-Sparse Quantization (ICML 2024 · similarity 0.88)
- ResQ: Mixed-Precision Quantization of Large Language Models with Low-Rank Residuals (ICML 2025 · similarity 0.88)
- RaZeR: Pushing the Limits of NVFP4 Quantization with Redundant Zero Remapping (NeurIPS 2026 · similarity 0.88)
- LeanQuant: Accurate and Scalable Large Language Model Quantization with Loss-error-aware Grid (ICLR 2025 · similarity 0.88)
- SpQR: A Sparse-Quantized Representation for Near-Lossless LLM Weight Compression (ICLR 2024 · similarity 0.88)
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.
