Step 3.7 Flash

Step 3.7 Flash was released by Shanghai-based StepFun at the end of May 2026 under the Apache 2.0 license. It is a sparse mixture-of-experts vision-language model with a 196B-parameter language backbone and a 1.8B vision encoder, activating roughly 11B parameters per token.

The model supports a 256K context window, three reasoning levels, and throughput up to 400 tokens per second per its model card, and StepFun pitches it as infrastructure for agents, coding work, search loops, and visual reasoning rather than a general chatbot. Performance claims still need real-world validation across hardware.