Breaking

StepFun's Step 3.7 Flash Ranks 8th Among Open-Weight Labs on Design Arena

June 21, 2026 at 21:42 EDT

Foundation Models
Open Source
Software Dev & Coding

StepFun, a China-based AI lab, has positioned its open-weight model "Step 3.7 Flash" among the top eight open-weight labs on Design Arena, a community-voted design-generation benchmark, trailing only Alibaba_Qwen and NexEcosystem. As an individual model it ranks 22nd among open weights with an Elo of 1216.

Late May 2026 · StepFun

Step 3.7 Flash: a 198B vision-language MoE built for coding agents

A sparse Mixture-of-Experts model that activates only ~11B parameters per token, adds native image understanding, and runs locally on 128 GB machines — released as open weights.

198B

Total params
(196B language + 1.8B ViT)

~11B

Active params
per token

256K

Context window

#22

Open-weight rank
Design Arena · Elo 1216

SWE-Bench Pro — Flash-tier coding

Higher is better. Step 3.7 Flash leads its closest open rival and closes on Gemini.

56.3%

Step 3.7
Flash

55.6%

DeepSeek
V4 Flash

55.1%

Gemini 3.5
Flash

Step-SWE-Bench: generation jump

67.08%

Step 3.7

56.5%

Prior gen

Speed & pricing

~385–400 tok/s in optimized setups
63–81 tok/s real-world API speeds
$0.20 input / $1.15 output per 1M tokens
Runs locally on 128 GB machines (~104–111 GB, Q4)

From Q&A toward agents

A "high-efficiency Flash model for real-world agents" — built around a perception → reasoning → action loop.

Perceive
screenshots · GUIs · docs · charts

→

Reason
low / medium / high

→

Act
coding · search · tool use

Praised for

Reliable agent workflows, low drift in long context
GUI understanding & screenshot-to-code
Punches above its ~11B active weight
Fast local inference on high-memory machines

Caveats

Verbose output
Full context / high precision needs lots of memory
Trails top closed frontier models on some agentic tasks

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.