MiniMax (official)· @MiniMax_AI · X·· 2026-07-14AI 评分42
AI 导读
在 RAISE 大会上，SambaNovaAI 将 H200 用于计算密集型预填充、SN50 RDUs 用于内存受限解码，实现了 MiniMax M2.7 的最快推理。经 Artificial Analysis 基准测试，短上下文达 850 t/s，长上下文超 450 t/s。
正文 · 原文
Paris brought the conversations. @SambaNovaAI brought the speed. 🇫🇷
at RAISE, @SambaNovaAI paired H200s for compute-heavy prefill with SN50 RDUs for memory-bound decode, reaching 850 t/s on short context and 450+ t/s on long context with MiniMax M2.7.
Worth reading the systems breakdown below👇
MiniMax_AI 2.7 just got even faster! At @RaiseSummit, SN50 delivered the fastest @MiniMax_AI M2.7 inference in the world, as benchmarked by @ArtificialAnlys. 850 t/s on short-context workloads. 450+ t/s on long-context workloads. Built for premium inference. Read more in our blog ⬇️ https://bit.ly/4fdcsrE在 X 查看被引用的帖子
来源：MiniMax (official) · x.com