😳震惊！vLLM直接砍掉Kimi K3.1 Pro 75%专家，体积暴降到476GB！
This is a AI post classified by AGK Intelligence as AI infra & evals (a tool drop), kept by the AI Radar because it carries real work, not commentary.
😳震惊！vLLM直接砍掉Kimi K3.1 Pro 75%专家，体积暴降到476GB！ 2.8万亿参数的超大MoE模型，被粗暴剪到只剩约700-750B参数，模型文件直接压到476GB，原始每层896个专家只保留224个，砍掉75%，属于校准无关的实验性剪枝，专门用来做vLLM服务和性能测试。 核心亮点： 🔹参数从2.8T暴降到约700-750B，体积只剩476GB 🔹每层专家：896 → 224（保留25%） 🔹定位是烟雾测试/性能调优，不是能正常用的模型 🔹GSM8K准确率从原版约95.5%直接掉到34.4%，质量回退非常明显 🔹官方明确说输出基本不可用，只适合跑服务和基准 使用步骤： 1️⃣安装支持该模型的vLLM（需trust-remote-code） 2️⃣直接运行： vllm serve mgoin/Kimi-K3-pruned75 \ --served-model-name kimi-k3-pruned-75pct \ --tensor-parallel-size 4 \ --trust-remote-code \ --gpu-memory-utilization 0.95 \ --max-model-len 32768 \ --kv-cache-dtype fp8 3️⃣服务起来后用OpenAI兼容接口调用测试 模型地址：http
Posted by SuSu_酥酥👅 (26.9k followers) 19 h ago · 27 likes · 18.3k views · view the original post on X. Kept by the AI Radar as AI infra & evals. Tools mentioned: kimi-k3-pruned75.
More AI work like this
- a16z's Martin Casado: AI is the biggest rebuild of computer science since the internet.… — @guilleflorvs
- There are AI GPUs running in converted chicken coops. — @guilleflorvs
- Enterprise AI is only as useful as the data and business context behind it. — @databricks
- DecodeWall: another benchmark I'm building, focused on concurrent calls. — @ivanfioravanti
- KindlingAI's GLM 5.3 Flash for one Spark recipe is here! 700k total KV on DGX OS (up to… — @mmastrac
- A small model coaching a giant? Aiona checks the held-out tie and SMF’s memory-card… — @MichaelGannotti
- Ranking inference providers six ways — @AdamHoltererer
- Harness adapters training is now live! — @yb2698
Every post is read and classified by AGK Intelligence: what it is, which market it belongs to, and whether the link is a real tool. 65.5k posts from 5k X accounts over the last 14 days, 10.4k tools, 19 markets. New posts are collected several times an hour and the full ranking is recomputed every hour — last update 2026-11-06 16:50 UTC. Full method.