Step 3.7 Flash

StepFun's Apache 2.0 vision-language agent model: 198B parameters with 11B active, a 256K context and a 111.5 GB GGUF for 128 GB machines.

In short. Step 3.7 Flash is an open model from the Chinese company StepFun that reads text and images and is built for agents that search, write code and use tools. It has about 200 billion parameters but uses 11 billion for each token, so it is fast for its size. A 4-bit file is 111.5 GB, which fits a Mac or mini PC with 128 GB of memory.

Summary

Step 3.7 Flash is StepFun's 198B-parameter sparse mixture-of-experts vision-language model, announced on 29 May 2026 under Apache 2.0; its Hugging Face repository first appeared on 23 May. It combines a 196B language backbone with a 1.8B vision encoder, activates about 11B parameters per token, has a 256K-token context and offers three selectable reasoning levels. StepFun claims up to 400 tokens per second and builds it for agentic work that mixes perception, search and reasoning, such as reading long financial reports or running concurrent coding agents. It is the successor of Step 3.5 Flash, with the same size and the addition of image input.

StepFun reports these results from its own table: 56.3 on SWE-Bench Pro (Step 3.5 Flash 51.3, DeepSeek V4 Flash 55.6, GPT 5.5 58.6, Claude Opus 4.7 64.3), 59.5 on Terminal-Bench 2.1 (DeepSeek V4 Flash 62.0), 49.5 on Toolathlon (DeepSeek V4 Flash 52.8), 45.8 on GDPval (Step 3.5 Flash 28.0) and 67.1 on ClawEval-1.1 (DeepSeek V4 Flash 57.8). On vision it reports 79.2 on SimpleVQA with tools and 95.3 on V* with Python, where Kimi K2.6 scores 96.9. StepFun itself points to Terminal-Bench and GDPval as areas where the model trails the best systems. All scores are StepFun's own.

The BF16 weights are 402.8 GB, with FP8 at 212.8 GB and NVFP4 at 129.0 GB. For local use StepFun publishes GGUF files, 111.5 GB at Q4_K_S, 105.0 GB at IQ4_XS, 102.5 GB at Q3_K_L, 209.4 GB at Q8_0 and 394 GB at BF16, plus a 4.0 GB vision projector. The card lists 120 GB of unified memory or VRAM as the minimum and 128 GB as recommended, on a Mac Studio, an NVIDIA DGX Station or an AMD Ryzen AI Max+ 395 system, and it points to StepFun's own llama.cpp branch and a StepFun vLLM Docker image. It also runs in SGLang and Transformers. In the 30 days to 11 November 2026, Hugging Face counted about 10,300 downloads of the main repository and about 5,900 of the GGUF files.

Good for

  • Local or on-premise agents on a 128 GB Mac or mini PC, using the 111.5 GB Q4_K_S file
  • Agents that read screenshots, charts and documents and then search, write code or call tools
  • Long documents and long agent runs, with a 256K context and selectable reasoning levels
  • Teams that want an Apache 2.0 vision-language model of about 200B in size

Watch out for

  • All scores are StepFun's own, and StepFun says it trails the best systems on Terminal-Bench and GDPval
  • Local use needs about 120 GB of memory and StepFun's own llama.cpp branch
  • The newer Step 5 Preview is larger and more capable by StepFun's account, but has a custom license
  • The 400 tokens per second figure is StepFun's claim for its own setup, not a typical laptop speed

Performance claims come from the developer and have not been independently verified.

How much memory does it need?

Estimates for the GGUF Q4_K_S variant. A mixture-of-experts model must hold all of its experts in memory even though only some are used for each token, so total parameters decide the memory and active parameters decide the speed.

Variants Parameters BF16 FP8 / INT8 4-bit (Q4) Fits on
Step-3.7-Flash (BF16) BF16 201B 11B active 463 GB 232 GB 139 GB Big-memory desktop
Step-3.7-Flash-FP8 FP8 201B 11B active 463 GB 232 GB 139 GB Big-memory desktop
Step-3.7-Flash-NVFP4 NVFP4 201B 11B active 463 GB 232 GB 142 GB Big-memory desktop
GGUF Q4_K_S INT4 201B 11B active 463 GB 232 GB 120 GB Big-memory desktop

Estimates: weights plus about 15% overhead. Real usage grows with context length and depends on the software. The developer states a minimum of 120 GB of GPU memory.

VRAM calculator

The license in plain words

One of the most widely used permissive licenses. You may use, change and sell products built on the model with only light conditions.

More about Apache 2.0

See also

Frequently asked questions

What is Step 3.7 Flash?
Step 3.7 Flash is an open model from the Chinese company StepFun that reads text and images and is built for agents that search, write code and use tools. It has about 200 billion parameters but uses 11 billion for each token, so it is fast for its size. A 4-bit file is 111.5 GB, which fits a Mac or mini PC with 128 GB of memory.
Can I use Step 3.7 Flash commercially?
Yes. Step 3.7 Flash is released under Apache 2.0, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
Is Step 3.7 Flash open source?
Step 3.7 Flash is released under Apache 2.0, a license approved by the Open Source Initiative. Training data and training code may not be public, so it is best described as open weights under an open-source license.
Can I run Step 3.7 Flash on my own computer?
Only on a machine with a lot of memory. The smallest variant, GGUF Q4_K_S, needs about 120 GB at 4-bit precision. Desktops with 128 to 512 GB of unified memory can hold it, at modest speed.
When was Step 3.7 Flash released?
Step 3.7 Flash was released by StepFun on May 29, 2026.
How long a context window does Step 3.7 Flash have?
Step 3.7 Flash supports a context window of up to 256K tokens.

Official links

Sources

Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.

Release date and facts last verified November 11, 2026.

Related models

Model Released Parameters Context License Smallest variant needs
Step 5 Preview StepFun 604B 27B active 1M StepFun 1.0 (Conditional) 417 GB Big-memory desktop
Step 3.5 Flash StepFun 199B 11B active 256K Apache 2.0 (Permissive) 120 GB Big-memory desktop
DeepSeek V4.1 Flash DeepSeek 552B 16B active 1M MIT (Permissive) 561 GB GPU server
Qwen3.8-27B Alibaba Qwen 27.8B 256K Apache 2.0 (Permissive) 19 GB Gaming GPU