81 GB smaller, not 81 GB. Don't plan on running this one at home. NVIDIA's NVF… · @qwerty__cs · TwKit

@qwerty__cs

81 GB smaller, not 81 GB. Don't plan on running this one at home. NVIDIA's NVFP4 build of DeepSeek-V4.1-Flash-1017 is still about 400 GiB, down from about 475 GiB. The routed experts were moved from MXFP4 to NVFP4 and the Engram tables from FP8 to MXFP4, while attention and the shared experts keep their original precision. The scores barely move. Terminal-Bench 2.1 goes from 81.81 to 81.11, SciCode from 57.58 to 57.32, and GPQA Diamond actually ticks up from 92.77 to 92.93. NVIDIA tested it on four GB300s. For now it only runs on vLLM, and you need a build with the MXFP4 Engram patch (PR #61530). MIT license. https://huggingface.co/nvidia/DeepSeek-V4.1-Flash-1017-NVFP4

曝光 86 · 评论 1 · 点赞 0 · 书签 0