minimax-m3.1POST
M3.1 Flash 权重上 HF：4 台 DGX Spark 可跑
作者给出本地部署观察：428B MoE、23B 激活、1M 上下文、多模态、约 250GB QAT 权重，Artificial Analysis 42 分与 GLM-5.3-Flash 持平；两台 DGX Spark 加上下文太紧，tp4 四台较宽裕，暂无 Spark 跑分。
𝙒𝙞𝙡𝙘𝙤 𝘷𝘢𝘯 𝘽𝙤𝙭𝙩𝙚𝙡 ⚡🚀🧲🌞
@nuonrg4 天前收录原推发布于 2026.10.23
𝙒𝙞𝙡𝙘𝙤 𝘷𝘢𝘯 𝘽𝙤𝙭𝙩𝙚𝙡 ⚡🚀🧲🌞 @nuonrg原推发布于 2026.10.23
MiniMax-M3.1-Flash open weights are live on Hugging Face: 428B MoE with 23B active, 1M context, multimodal, ~250GB QAT checkpoint. Artificial Analysis has it at 42, on par with GLM-5.3-Flash. Too tight for 2 DGX Sparks once you add context, but it fits comfortably on 4 with tp4. No Spark benchmarks yet. https://huggingface.co/MiniMaxAI/MiniMax-M3.1-Flash
x.com/nuonrg/status/2113538680807625091采集于 2026.10.24 12:10