deepseek-v4.1POST
V4.1-Flash-1017 首个 NVFP4 量化版上线 Hugging Face，约 430GB
专家权重由 MXFP4 无损转码为 NVFP4，约 430GB，可放进 4 台 DGX Spark 并留约 80GB 给 KV cache；作者称 2 台 Spark 的版本在开发中。
𝙒𝙞𝙡𝙘𝙤 𝘷𝘢𝘯 𝘽𝙤𝙭𝙩𝙚𝙡 ⚡🚀🧲🌞
@nuonrg10 天前收录原推发布于 2026.10.17
𝙒𝙞𝙡𝙘𝙤 𝘷𝘢𝘯 𝘽𝙤𝙭𝙩𝙚𝙡 ⚡🚀🧲🌞 @nuonrg原推发布于 2026.10.17
First NVFP4 build of DeepSeek V4.1 Flash 1017 is already on Hugging Face: lossless MXFP4 to NVFP4 transcode of the experts, about 430GB. That fits on 4 DGX Sparks with roughly 80GB left for KV cache. A 2-Spark build is reportedly in the works. GLM-5.3 Flash finally has competition on my cluster. #AI #LocalAI #Deepseek https://huggingface.co/LibertAIDAI/DeepSeek-V4.1-Flash-1017-NVFP4
x.com/nuonrg/status/2111393037406359586采集于 2026.10.18 16:45