StepFun releases Step 3.7 Flash, a 198B vision-language MoE with 11B active parameters

StepFun published Step 3.7 Flash. The card describes a 198B-parameter sparse mixture-of-experts vision-language model: a 196B language backbone plus a 1.8B vision encoder, with about 11B parameters active per token. It claims up to 400 tokens per second, a 256k context window and three reasoning levels (low, medium, high).

The card targets agentic work that combines perception, search and reasoning, such as parsing long financial reports or running concurrent coding agents. It reports 56.3 on SWE-Bench Pro, 59.5 on Terminal-Bench 2.1 and 67.1 on ClawEval-1.1. All are StepFun's own figures, and it says Terminal-Bench and GDPVal-AA show areas for future work.

The licence is Apache 2.0.

The card gives no release date; the date used is the Hugging Face repo creation date.

A card that admits where it trails. Take the leading claims with the same grain of salt.

Sources

• Model card on Hugging Face
• Hugging Face repo metadata (creation date, licence, size)

Editor's first pass (AI-assisted)

• Real 4: Apache 2.0, ungated, with FP8, NVFP4 and GGUF variants listed in the repository.
• Useful 3.5: Aimed at agent loops that mix images, search and coding at high throughput.
• Notable 3.5: 462 likes; a mid-sized entrant in a crowded field.
• Novel 3: Three selectable reasoning levels and a vision encoder on a 196B backbone.