AI Paper Atlas

Research map / Architectures and efficiency / Quantization

To Compress or Not? Pushing the Frontier of Lossless GenAI Model Weights Compression with Exponent Concentration

Zeyu Yang, Tianyi Zhang, Jianwen Xie, Chuan Li, Zhaozhuo Xu, Anshumali Shrivastava

ICLR 2026 · Architectures and efficiency › Quantization · approach: Low-precision training formats

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
losslessly compress FP8 weights with ECF8 (Huffman-coded exponents)FP8 LLMs · Weight memorymemory reduction: 9.8% to 14.8%Improved
losslessly compress FP8 weights with ECF8FP8 DiTs · Weight memorymemory reduction: 14.1% to 26.9%Improved
use ECF8 memory savings to enlarge batch size under fixed memoryDeepSeek-R1-0528 · 640 GB budgetthroughput: 1.55 -> 3.88 tokens/s (150.3%)Improved
use ECF8 memory savings to enlarge batch size under fixed memoryDeepSeek-R1-0528 · 640 GB budgetper-request latency: 660.65 -> 263.95 sImproved
use ECF8 under fixed memoryLlama-3.3-70B-Instruct-FP8-dynamic · 80 GB budgetthroughput: 41.28 -> 45.96 tokens/sImproved
use ECF8 under fixed memoryQwen3-8B-FP8 · 12 GB budgetthroughput: 208.80 -> 235.22 tokens/sImproved
use ECF8 with DiffSynth VRAM managementFLUX.1-dev · DiffSynth GH200E2E latency: 24.29 -> 13.15 sImproved
use ECF8 with DiffSynth VRAM managementQwen-Image · DiffSynth GH200E2E latency: 111.14 -> 49.05 sImproved
use ECF8 with DiffSynth VRAM managementWan2.1-T2V-14B · DiffSynth GH200E2E latency: 476.21 -> 460.67 sImproved
use ECF8 with DiffSynth VRAM managementWan2.2-T2V-A14B · DiffSynth GH200E2E latency: 480.45 -> 461.41 sImproved
ECF8 compression on generated outputsQwen-Image · Generated images, fixed seedoutput fidelity: pixel-identicalNo clear effect

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