Chinese AI startup StepFun has open-sourced Step 3.7 Flash, a new generation of its Flash series large language model designed for production-grade agents, coding, web search, and multimodal workflows. The model is released under the Apache 2.0 license and uses a sparse MoE architecture with a 196B language backbone paired with a 1.8B vision transformer (ViT), totaling 198B parameters but activating only 11B. It supports a 256K context window, three inference tiers, and a peak speed of 400 tokens per second. StepFun says it has optimized the model for local hardware including the Mac Studio M4 Max, DGX Spark, and AMD AI Max + 395.
Positioned as a foundation for production-grade agents, Step 3.7 Flash boasts native multimodal understanding and execution capabilities. It can autonomously select, crop, and reread UI elements or charts, and it proactively searches for verification when information is uncertain. The model topped the ClawEval-1.1 autonomous task execution benchmark with 67.1% accuracy, and showed strong engineering reasoning on SimpleVQA Search (79.2%), V* Python (95.3%), and SWE-Bench Pro (56.3%, ranking second globally). In multi-tool coordination and task execution, it scored 49.5% on Toolathlon, 45.8% on GDPval (covering 44 professions), and achieved over 98% pass rates across all difficulty levels of the τ²-bench communication benchmark.
StepFun has ensured deep compatibility with agent frameworks like Claude Code, OpenClaw, KiloCode, RooCode, and Hermes Agent, as well as the MCP protocol. The company also released open-source interfaces on Hugging Face and ModelScope, a GGUF version for edge deployment, and completed integrations with platforms including OpenRouter, ZenMux, and Fireworks AI.