Step 3.7 Flash

What is Step 3.7 Flash?

Step 3.7 Flash is a mixture-of-experts model with a 196B-parameter language backbone and a 1.8B-parameter vision encoder, and only about 11B parameters work on each token. It reads text, images and short videos, holds 256K tokens of context and calls tools through an OpenAI-compatible API, with low, medium and high reasoning settings.

It builds on text-only Step 3.5 Flash, which StepFun still sells, and costs twice as much for input and nearly four times as much for output. StepFun's larger Step 5 Preview is now its flagship. The weights are free under Apache 2.0.

What changed from Step 3.5 Flash?

Step 3.7 Flash adds image and video input and a choice of reasoning effort, and scores higher on StepFun's coding and agent tests, at a higher token price.

Input

A 1.8B-parameter vision encoder adds image and video understanding; Step 3.5 Flash reads text only.

Reasoning

Three effort levels, low, medium and high, with medium as the default.

Scores

StepFun reports 56.3% on SWE-Bench Pro against 51.3%, and 67.1% on ClawEval-1.1 against 43.6%.

Price

$0.20 / $1.15 per million tokens against $0.10 / $0.30, so output costs nearly four times as much.

Unchanged

Both activate about 11B parameters per token, hold 256K tokens and ship under Apache 2.0.

Who is Step 3.7 Flash best for?

Step 3.7 Flash is best for developers building agents that read screens, documents or video and call tools at high volume, through StepFun's API or on a 128 GB machine of their own.

• Agents that read screens and documents

  It takes several images per request plus MP4 video, so an agent can read interfaces, charts and scanned pages, then act on them with tools.

• Coding agents in common harnesses

  StepFun reports 56.3% on SWE-Bench Pro and documents setups for Claude Code, OpenClaw, Cline, Roo Code and Kilo Code.

• Search-heavy research

  StepFun trained it to plan searches and check sources rather than rely on memory, and reports 75.8% on BrowseComp.

• Running a vision model yourself

  The 4-bit build and its vision projector fit a 128 GB Mac Studio, NVIDIA DGX Station or AMD Ryzen AI Max+ 395 machine.

What's included with Step 3.7 Flash?

Image and video input, three reasoning levels, Apache 2.0 weights with FP8, NVFP4 and GGUF builds, support in vLLM, SGLang and llama.cpp, and a Step Plan for coding tools.

Image and video input

JPG, PNG, WebP and static GIF images by URL or Base64, several per request, plus a single MP4 under 128 MB by URL.

Reasoning effort

Low, medium or high per request through reasoning_effort, with medium as the default.

Open weights

Apache 2.0 weights on Hugging Face and ModelScope, with FP8, NVFP4 and GGUF builds.

Runs on common servers

StepFun documents vLLM, SGLang, Hugging Face Transformers and llama.cpp, and NVIDIA offers it as a NIM microservice.

Step Plan for coding tools

Monthly credits that work in Claude Code, Cursor, OpenClaw and Trae, from $6.99 a month.

Phone agent guide

A StepFun guide connects it to a real Android phone, where it plans each action from screenshots.

How much does Step 3.7 Flash cost?

Step 3.7 Flash costs $0.20 per million input tokens and $1.15 per million output tokens on StepFun's API, Step Plan subscriptions start at $6.99 a month, and the weights are free to download.

Flash Max costs $99 a month for 40,000M credits. Paid quarterly, Mini, Plus, Pro and Max cost $18.99, $26.99, $79 and $269; paid yearly, $69.99, $95.99, $289 and $989. StepFun values about 7M credits at $1 of model use; credits arrive monthly, expire at month end, and booster packs cost $6.99 for 400M or $9.99 for 1,600M. The weights are free under the Apache 2.0 licence on Hugging Face and ModelScope. Full pricing on platform.stepfun.ai. We checked these prices on 29 Sep 2026.

How do readers rate Step 3.7 Flash?

No one has rated Step 3.7 Flash yet. Be the first.

Step 3.7 Flash pros and cons

It adds vision to a fast, cheap open model and some users rate it above Qwen 3.6 27B for coding, but it reasons at length, can be uneven and costs more per token than Step 3.5 Flash.

Is Step 3.7 Flash open source?

Yes. Step 3.7 Flash's code is public under the Apache 2.0 licence, so you can run it on your own server.

What are the best Step 3.7 Flash alternatives?

Step-3.5-Flash for cheaper text-only work, DeepSeek V4.1 Flash for a 1M context, and Qwen3.8-27B for a vision model that needs far less memory.

Step-3.5-FlashOpen Source

Choose it when your agents only handle text; it costs $0.10 / $0.30 per million tokens, has the same 256K context and is in the same Step Plan.

DeepSeek V4.1 FlashOpen Source

Choose it when you need a 1M-token context with image input on the API; it costs $0.30 / $1.20 per million tokens at peak hours and half that off-peak, with MIT-licensed weights.

Qwen3.8-27BOpen Source

Choose it when you want an open vision model that needs far less memory: a 27B dense model under Apache 2.0 that reads images and video, with a 262K native context.

Questions about Step 3.7 Flash