Step 3.7 Flash 把推理效率从模型设计阶段就考虑进去，做智能体或搜索应用的团队可以直接在 Fireworks 上试用，成本可能比 DeepSeek 低很多。
StepFun 发布 Step 3.7 Flash，一个 196B 参数的 MoE 模型，从设计之初就针对推理效率进行优化。它采用多矩阵分解注意力（MFA）技术，KV-cache 仅为 DeepSeek 的约 22%，并通过注意力-FFN 解耦（AFD）实现硬件优化的服务。模型在 ClawEval-1.1、SimpleVQA Search 等基准测试中排名第一，支持 400 TPS 的推理速度，256K 上下文窗口，并具备三种推理级别。它专为智能体、编程、搜索和多模态工作流设计，支持本地运行（如 Mac Studio M4 Max），并采用 Apache 2.0 许可证开放权重。
Many research labs only consider inference efficiency after the fact. Step 3.7 Flash is a 196B MoE m...
Many research labs only consider inference efficiency after the fact. Step 3.7 Flash is a 196B MoE model, and built for inference from the start by @StepFun_ai . Multi-Matrix Factorization Attention (MFA) → KV-cache at ~22% of DeepSeek. Attention-FFN Disaggregation (AFD) → decoupled attention + FFN for hardware-optimized serving. Try it now on Fireworks. Apache 2.0 license → fireworks.ai/models/firewor… StepFun @StepFun_ai ⚡️ Step 3.7 Flash is here: The new frontier is agent efficiency. #1 ClawEval-1.1 (67.1), #1 SimpleVQA Search (79.2), #2 SWE-PRO (56.3), 95.3 on V* Python. Open weights under Apache 2.0. Built for agentic, coding, search, and multimodal workflows — balancing speed, cost, and reliable execution. - 400 TPS. 198B sparse MoE, ~11B active. 256K context, 3 reasoning levels. - Understands UIs, charts, docs, images — then writes code or calls tools to act on what it sees. - Web + visual search reaches further: more sources, deeper follow-up. - Reliable tool use — less drift, fewer broken toolcalls. 98%+ on τ²-bench across all difficulty levels. - Works with Claude Code, KiloCode, Hermes Agent, OpenClaw, and protocols like MCP. - Runs locally on Mac Studio M4 Max, DGX Spark, AMD AI Max+ 395. GitHub: github.com/stepfun-ai/Ste… HuggingFace: huggingface.co/stepfun-ai/Ste… GGUF: huggingface.co/stepfun-ai/Ste… ModelScope: modelscope.cn/models/stepfun… API: platform.stepfun.ai Blog: static.stepfun.com/blog/step-3.7-… 🔗 View Quoted Tweet 💬 0 🔄 1 ❤️ 1 👀 149 📊 1 ⚡