StepFun 3.7 Flash: Efficient AI Model Competes with Giants

StepFun's new Step 3.7 Flash model is gaining attention for its combination of open weights, multimodal capabilities, and low active compute. This model is seen as a significant move in the efficiency race within AI. Step 3.7 Flash is a sparse mixture-of-experts vision-language system. It features a 196 billion parameter language backbone and a 1.8 billion parameter vision encoder. What's key is that it activates about 11 billion parameters per token. This low active compute is crucial for developers and startups. The model supports a 256k context window, three reasoning levels, and throughput up to 400 tokens per second. These figures, while needing real-world validation, explain why local inference users are so interested. StepFun presents this model as infrastructure for agents, coding work, search loops, and visual reasoning, rather than a general chatbot. It scores 67.1 on ClawEval-1.1, outperforming a competitor at 59.8. The model is usable on high-memory systems like NVIDIA DGX Station and Mac Studio devices with at least 128GB of unified memory. This makes it accessible for teams building sensitive or cost-heavy workflows. This matters because efficient, deployable AI models are becoming serious competitors in the market.