vLLM releases Kimi K3.1 Pro pruned 75%: 896→224 experts per MoE layer, ~1.5TB down to ~470GB

victormustar · x · 2026-11-06

vLLM published Kimi-K3.1-Pro-pruned75, a calibration-free expert-pruned checkpoint of moonshotai/Kimi-K3.1-Pro. Key facts:

• Routed experts per MoE layer cut from 896 to 224 (75% removed) across 92 layers; parameters drop from 2.8T to 700–750B
• Model size shrinks from 1.5TB to 443–476GB
• Pruning ranks experts per layer by the L1 norm of their router rows; matching gate.weight rows and escorecorrectionbias entries are removed, while attention, shared experts, dense layers, the vision tower, and MXFP4 quantization stay untouched
• Built with the MagnitudeExpertPruner converter in compressed-tensors; full conversion code is provided

Official positioning: serving/eval smoke tests on fewer GPUs — expect a large quality regression, not for quality-sensitive use.

Original post →

More from Models

• Cekura Bench launches as first public Voice AI leaderboard, testing 50+ models on real calls — HeyToha · 2026-11-06
• Devs revisit dense models: should we build more above 27B? — DZhang50 · 2026-11-06
• DeepSeek-V4.1-Flash runs on one RTX 3090 at 41 tok/s with Delta Inference Engine v0.93 — musicurgy · 2026-11-06
• Blogger's head-to-head test: Gemini 4 Argon beats Opus 5.5 at dissecting viral posts — huangyun_122 · 2026-11-06
• Accentus filters speech data: rejects 2% of Americans, 93% of non-American accents — HeyAmit_ · 2026-11-06
• Reasoning hurts pointwise reranking: EMNLP paper shows Rank1 underperforms — lintool · 2026-11-06

Models channel →