Small Is Beautiful: StepFun Drops 198B Open-Source MoE Model — 11B Active, 400 Token/s, Agent Community Erupts

StepFun releases Step 3.7 Flash: 198B MoE multimodal open-source model, only 11B active params, 400 token/s throughput, compatible with OpenAI & Anthropic protocols. Agent cost king.

💡 What You Will Learn

StepFun releases Step 3.7 Flash: 198B MoE multimodal open-source model, only 11B active params, 400 token/s throughput, compatible with OpenAI & Anthropic protocols. Agent cost king.

On May 29, StepFun quietly posted on X with a rocket-launch image. Click through, and there it is: Step 3.7 Flash — open-source, multimodal, Apache 2.0 licensed. This isn't just another large chatbot. It's a compact powerhouse purpose-built for agent production environments — 198B total parameters, but only 11B activated per inference, with generation speeds hitting 400 tokens/s. To put that in perspective: by the time you finish reading this sentence, it's already generated half a page of text. Three numbers tell the story. 198B → 11B: a sparse MoE architecture with a large total parameter count but only 11B activated per token. This means low barriers to entry — inference runs on consumer-grade GPUs without needing racks of H100s. 400 tokens/s: currently the fastest in its class among open-source multimodal models. The previous speed leader, GPT-5.3, managed 100+ token/s. Step 3.7 Flash more than triples that. For agent scenarios where a single task might invoke the model dozens of times, every millisecond compounds into a transformative user experience.