minimax-m3.1POST
MiniMax M3.1 Flash weights: MoE architecture and official scores
A rundown of the newly released M3.1 Flash weights: a 428B-total, 23B-active MoE with sparse attention and DSpark speculative decoding, official scores of 73.7% on DeepSWE v1.1 and 82.2 on CyberGym, with community quants and vLLM throughput still to come.
JL
@LJL06019Archived 4 days agoOriginally posted 2026.10.23
JL @LJL06019Originally posted 2026.10.23
开了。昨晚蹲的 MiniMax M3.1 Flash,权重刚放出来。 新料主要是架构:总参数 428B,每 token 只激活 23B——MoE。配上它自家的稀疏注意力和 DSpark 推测解码,冲着"大个子、低成本"去的。 分数还是那套(官方自报):DeepSWE v1.1 73.7%、OSWorld V2 61.7、CyberGym 82.2,跟 API 发布时给的 74% 基本对得上。 接下来蹲社区量化版和 vLLM 这些推理框架的实测吞吐。428B 全量是服务器的活,但激活只有 23B,这笔账最后还是算在推理成本上——Flash 这一档,拼的就是这个。
This post quotes another post: “MiniMax M3.1 Flash open weights: 428B total, 23B activated”
x.com/ljl06019/status/2113443219858161965Archived 2026.10.24 12:10