Stepfun · 2026-05-29 · major
Stepfun Releases Step 3.7 Flash — 198B Vision-Language MoE With 11B Active Parameters, 56.3% SWE-Bench Pro, 256K Context, and Three Reasoning Levels
Sparse MoE vision-language flagship: 196B language backbone + 1.8B vision encoder, ~11B active per token, 256K context, up to 400 tok/s, and three selectable reasoning levels. Apache 2.0 with weights on Hugging Face.
198B-parameter vision-language MoE tuned for high-frequency agentic workloads — Apache 2.0, 256K context, and three reasoning levels you can dial up or down.
Key specs
| Parameters | 196B + 1.8B ViT |
|---|---|
| Active params | 11B |
| Context window | 256K |
| Throughput | up to 400 tok/s |
| Swe bench pro | 56.3 |
| Terminal bench 2 1 | 59.6 |
| Toolathlon | 49.5 |
| Simplevqa | 79.2 |
| V star with python | 95.3 |
What is it?
Step 3.7 Flash is Stepfun's new multimodal model in the Flash line. It is a sparse mixture-of-experts that combines a 196B language backbone with a 1.8B vision encoder, activating about 11B parameters per token for production-grade throughput.
How does it work?
The architecture pairs MoE routing with a native vision encoder, so images flow into the same context as text without a separate adapter. Inference exposes three reasoning levels (low / medium / high) plus an Advisor Mode that picks the cheapest level capable of solving the task. Visual search, GUI operation, and tool calls share the same compositional pipeline, which is what lifts SWE-Bench Pro from Step 3.5's 51% range to 56.3%.
Why does it matter?
Open-weights vision-language MoEs that hit 50%+ on SWE-Bench Pro are still rare, and Apache 2.0 weights with 256K context let teams run agentic coding and multimodal workflows on their own infrastructure. Step 3.7 Flash gives Stepfun a credible Flash-tier answer to GPT-5.5 / Claude Opus 4.x for cost-sensitive deployments.
Who is it for?
agent builders deploying multimodal pipelines on their own GPUs
Try it
huggingface-cli download stepfun-ai/Step-3.7-Flash