deepseek-v4.1POST
V4.1-Flash-1017 在双 DGX Spark 上 EXL3 2.9bpw 量化可跑
约 210GB 量化权重，代码困惑度仅损失约 2%，可直接套用现有 V4.1 Spark 套件，但需换新 1017 Engram 表。
𝙒𝙞𝙡𝙘𝙤 𝘷𝘢𝘯 𝘽𝙤𝙭𝙩𝙚𝙡 ⚡🚀🧲🌞
@nuonrg11 天前收录原推发布于 2026.10.17
𝙒𝙞𝙡𝙘𝙤 𝘷𝘢𝘯 𝘽𝙤𝙭𝙩𝙚𝙡 ⚡🚀🧲🌞 @nuonrg原推发布于 2026.10.17
DeepSeek V4.1 Flash 1017 is now runnable on 2x DGX Spark in EXL3: a 2.9 bpw quant (~210 GB) with only ~2% perplexity loss on code, drop-in for the existing V4.1 Spark kit. Remember to pull the new 1017 Engram tables, the old ones changed. #AI #Deepseek #LocalAI https://huggingface.co/LibertAIDAI/DeepSeek-V4.1-Flash-1017-EXL3-2.9bpw
x.com/nuonrg/status/2111473990065590666采集于 2026.10.18 16:45