Step 3.7 Flash

What it is

Step 3.7 Flash is a sparse Mixture-of-Experts vision-language model from StepFun (阶跃星辰). It combines a 196-billion-parameter language backbone with a 1.8-billion-parameter vision encoder, and about 11 billion parameters are active at a time.^[2] A vision encoder turns an image into a form the language model can work with. A vision-language model is one that accepts images as well as text.

StepFun released Step 3.7 Flash on 29 May 2026 as a high-efficiency model for real-world agents, meaning AI systems that carry out tasks on a user's behalf.^[1]

Origins

Step 3.7 Flash builds on the design of Step 3.5 Flash. The main change is image understanding: the earlier model lacked native image input, and Step 3.7 Flash adds it.^[1][2] The model can also use Python-based visual tools, such as cropping and zooming an image, so that it can look more closely at part of a picture before answering.^[1]

Main features

The model supports a 256K context and offers low, medium and high reasoning levels, letting users trade speed against depth of reasoning. StepFun says it can generate up to 400 tokens per second.^[2]

Pricing for the API was US$0.20 per million input tokens when the input was not in the cache, US$0.04 per million when it was a cache hit, and US$1.15 per million output tokens.^[2] A cache hit means the system reuses input it has already processed, which is cheaper. The model is available on the StepFun Open Platform, OpenRouter and NVIDIA NIM.^[2] The weights are released under the Apache 2.0 license.^[2][3]

StepFun described an "Advisor Mode". In this mode the model runs most of a task itself and escalates to a larger model at key points.^[1] StepFun said that, using this mode, the model reached 97% of Claude Opus 4.6's coding performance at about one-ninth of the per-task cost, $0.19 against $1.76.^[1] StepFun also said the model works with agent harnesses, which are the programs that wrap a model and let it act, including Claude Code, KiloCode, Hermes Agent and OpenClaw.^[1]

Reported results

StepFun reported the following scores, with the Step 3.5 Flash result in brackets where it was given:

• SWE-Bench Pro: 56.3% (51.3% for Step 3.5 Flash). This benchmark tests fixing real software issues in a harder form.^[1]
• SWE-bench Verified: 76.5%.^[1]
• Toolathlon: 49.5% (33.3% for Step 3.5 Flash). This benchmark tests a model's ability to use many tools.^[1]
• BrowseComp: 75.8% (69.0% for Step 3.5 Flash). This benchmark tests finding hard-to-locate information on the web.^[1]
• HLE with tools: 47.2% (35.7% for Step 3.5 Flash). HLE is a difficult set of expert-level questions, here allowed to use tools.^[1]

Where it stands

Step 3.7 Flash's stated focus is efficiency in agent work, as its release tagline suggests. The company's own comparisons are the main source for the performance claims, so readers should check them against independent tests. The open weights and the Apache 2.0 license make the model available for others to examine and run, within the terms of the license.