Step 3.7 Flash
198B vision-language MoE (11B active, 256K context) with three reasoning levels; 56.3 on SWE-Bench Pro and first place on ClawEval at 67.1.
Adds native image understanding (1.8B vision encoder) to the Flash line and up to 400 tokens/s. Priced at $0.20 input, $0.04 cached, $1.15 output per million tokens. Apache-2.0; weights on HF from 2026-05-23.
- Date
- Friday, 29 May 2026
- Lab
- StepFun
- Kind
- open-weights
- Access
- open weights
- Price
- $0.20 input / $1.15 output per M tokens, 2026-05
Figures
| Measure | Value | Measured by |
|---|---|---|
| SWE-Bench Pro | 56.3 vs Step 3.5 Flash 51.3 | company |
| Terminal-Bench 2.1 | 59.5 vs Gemini 3.5 Flash 76.2 on same chart | company |
| ClawEval-1.1 | 67.1 next best 59.8 per StepFun | company |
Blog dated 2026-05-29. Benchmarks are company-run; StepFun's own chart shows it behind Gemini 3.5 Flash and GPT-5.5 on Terminal-Bench 2.1.
Sources
This record was checked against its sources on 6 October 2026. How we check