小米 MiMo 开源 MiMo-V2.5-Pro-UltraSpeed，1T 模型单节点突破 1000 tps
🔓 And the best part — we're open-sourcing it.
小米 MiMo 与 TileRT 联合发布 MiMo-V2.5-Pro-UltraSpeed，在单个标准 8-GPU 节点上让 1T MoE 模型输出速度突破 1000 tokens/s，并开源 FP4 + DFlash 检查点权重。该速度来自模型与系统的协同设计：模型侧采用 FP4 量化与 DFlash 块掩码并行推测解码，系统侧 TileRT 针对这些技术定制编译器与 kernel，全程使用通用 GPU，未采用 Cerebras 晶圆级或 Groq SRAM ASIC 方案。UltraSpeed 提供限时免费网页聊天，API 价格约为 3 倍换取约 10 倍输出体验，申请通道开放至 6 月 23 日（PDT）。
正文
当前显示原文，中文全文译文尚未就绪。
🔓 And the best part — we're open-sourcing it.
1,000+ tps on a 1T model wasn't a single breakthrough — it's deep model × system co-design between the MiMo and TileRT teams, all on general-purpose GPUs (no Cerebras-style wafer-scale, no Groq-style SRAM ASICs). On the model side: FP4 quantization (smaller footprint, less memory traffic) + DFlash, our block-masked parallel speculative decoding that accepts far more tokens per verification. On the system side, TileRT tailors its compiler & kernels to exactly these techniques.
The result: a 1T model breaking 1,000 tps on a single, standard 8-GPU node.
🤗 Open weights (FP4 + DFlash checkpoint): https://t.co/jYQsgeruMg