概念

sparse MoE

Q: sparse MoE 最近有什么新动态？

traeai 已收录 1 篇与 sparse MoE 相关的内容。最新一篇是「Many research labs only consider inference efficiency after the fact. Step 3.7 Flash is a 198B spars...」，由 Fireworks AI(@FireworksAI_HQ) 发布。

别名：稀疏MoE

稀疏专家混合架构，用于在保持模型能力的同时显著降低推理资源消耗。

已跟踪 1 条高相关材料

TraeAI 观察

如果只读 3 篇

Many research labs only consider inference efficiency after the fact. Step 3.7 Flash is a 198B spars...

Fireworks AI(@FireworksAI_HQ) · 8.2 分

Fireworks AI 引入 Step 3.7 Flash：198B 稀疏 MoE 视觉语言模型，从推理出发设计，集成 196B 语言与 1.8B 视觉编码器，实现实际代理任务最高 400 token/s 推理。

Fireworks AI on X: "Many research labs only consider inference efficiency after the fact. Step 3.7 Flash is a 198B sparse MoE VLM designed by @StepFun_ai for inference from the start. 196B language backbone with a 1.8B vision encoder. Built for real-world agent workloads, running at up to 400 ..."

Fireworks AI(@FireworksAI_HQ)6月4日112 字 (约 1 分钟)

Fireworks AI introduces Step 3.7 Flash: a 198B sparse MoE VLM designed for inference from the start, with a 196B language backbone and 1.8B vision encoder, achieving up to 400 token/s on real-world agent workloads.

入选理由：从设计阶段即优化推理效率，非事后补强。

FeaturedTweet#Step3.7 Flash#sparse MoE#VLM#198B#400 token/s英文

跨材料问答 · sparse MoE

回答基于：sparse MoE 相关 1 条材料