Step 3.7 Flash
Editorial notes
Lanzado 28 mayo 2026. Nuevo flagship multimodal (Step 3.5 Flash era text-only). 198B sparse MoE / ~11B activos + 1.8B ViT vision encoder. Hasta ~400 tok/s. Apache 2.0, open-source en GitHub + HF + ModelScope. API en StepFun Platform, OpenRouter, NVIDIA NIM, DeepInfra, Fireworks. Pricing $0.20/$1.15 per MTok (cache hit $0.04 input). GPQA-Diamond 77.8 medido con NVFP4+MTP. Otros scores del repo oficial stepfun-ai/Step-3.7-Flash: ClawEval-1.1 67.1, Terminal-Bench 2.1 59.5, Toolathlon 49.5, HLE con tools 48.1, GDPVal-AA 45.8, SimpleVQA 79.2, V* 95.3.
Spec sheet
- Company
- StepFun
- Country
- CN
- Type
- multimodal
- Release
- 2026-05
- Context
- 262K tokens
- Params total
- 198B
- Params active (MoE)
- 11B
- License
- apache-2.0
- Quants
- BF16, Q8_0, Q5_K_M, Q4_K_M
- Pricing (openrouter)
- $0.2/$1.15/M
- Slug
- step-3-7-flash
Quick install
1 toolsvllm serve stepfun-ai/Step-3.7-FlashBenchmarks (2)
Cite this model
BibTeX · APA
BibTeX
@misc{frontier-step-3-7-flash,
title = {Step 3.7 Flash},
author = {{StepFun}},
year = {2026},
note = {Frontier Benchmarks AI atlas. Accessed 2026-06-10},
url = {https://frontierbenchmarks.com/models/step-3-7-flash}
} APA
StepFun (2026). Step 3.7 Flash [Large language model]. Frontier Benchmarks AI. Retrieved 2026-06-10, from https://frontierbenchmarks.com/models/step-3-7-flash
Citation reflects the atlas page, not the original model paper. For the paper, see the "Resources" section above.