Step 3.7 Flash: Apache 2.0 multimodal MoE tuned for agent efficiency

Per StepFun's launch post, Step 3.7 Flash is a 198-billion-parameter sparse mixture-of-experts model that activates roughly 11 billion parameters per token, with a 256K-token context window, three selectable reasoning levels, and open weights under the Apache 2.0 license. In a sparse mixture-of-experts model, each token is routed through only a small subset of specialized sub-networks, called experts, instead of running the model's entire parameter set on every input, letting it behave like a much larger model while computing through a fraction of its total parameters per step.

StepFun reports the model ranking first on two benchmarks it names, ClawEval-1.1 at 67.1 and SimpleVQA Search at 79.2, second on SWE-PRO at 56.3, and 95.3 on a Python-focused test it calls V* Python, alongside throughput of 400 tokens per second and tool-call reliability above 98 percent on tau-squared-bench across all difficulty levels. These are the company's own self-reported figures at launch, not independently reproduced here. StepFun also says the model works with existing agent tools including Claude Code, Kilo Code and the Model Context Protocol, and can run locally on hardware like a Mac Studio M4 Max or Nvidia DGX Spark.