StepFun's Step 3.7 Flash Hits Frontier Agentic Performance at One-Ninth the Cost
StepFun's 198B MoE model delivers frontier-class agentic coding and vision at roughly one-ninth the cost of Claude Opus 4.6
Takeaways
- What: StepFun released Step 3.7 Flash, a 198B sparse MoE vision-language model activating only ~11B parameters per token.
- Cost: $0.20/M input tokens and $1.15/M output tokens; runs agentic coding tasks at ~$0.19 vs. ~$1.76 for Claude Opus 4.6.
- Vision-first: First multimodal model in StepFun's Flash line, adding a 1.8B ViT encoder; Step 3.5 Flash was text-only.
- Benchmarks: #1 on ClawEval-1.1 (67.1), #1 on SimpleVQA Search (79.2), 56.3% on SWE-Bench Pro; lags on Terminal-Bench (59.5 vs. 76.2 for Gemini 3.5 Flash).
- Deployment: Apache 2.0 weights in BF16/FP8/NVFP4/GGUF on Hugging Face; runs locally on 128GB devices (Mac Studio, DGX Spark); also on OpenRouter and NVIDIA NIM.
- Architecture: Hybrid sliding-window + global attention, 288 experts per layer with top-8 routing, and built-in MTP-3 speculative decoding for up to 400 tokens/sec.
Step 3.7 Flash is the new open-weight model from Chinese AI lab StepFun, and it is making a pointed argument: you do not need a trillion-parameter behemoth to run production agentic workflows. With 198 billion total parameters but only 11 billion active per token, it is engineered to sit at the intersection of frontier reasoning and Flash-tier cost , and its benchmarks largely back that claim up.
A model built for loops, not leaderboards
Step 3.7 Flash is an open-weight agentic vision model built on a 198B-parameter sparse Mixture-of-Experts (MoE) architecture that activates only about 11B parameters per token. It pairs a 196B language backbone with a 1.8B vision encoder, ships under Apache 2.0, and is aimed squarely at high-frequency production agentic workflows rather than leaderboard supremacy. The practical consequence: a model that runs roughly 11B parameters of compute per token, prices output at $1.15 per million tokens, and runs an agentic task for about $0.19 where a frontier model runs the same task for closer to $1.76.
It is the third entry in StepFun's Step-3 family: Step-3 (July 2025), the text-only Step-3.5-Flash (February 2026), and now Step-3.7-Flash, the first multimodal model in the Flash line. Step 3.5 Flash had no multimodal support; this is a new addition in 3.7.
The architecture under the hood
In a standard dense model, every parameter fires for every token. MoE (Mixture-of-Experts) breaks that by dividing the network into many specialized sub-networks called "experts" and routing each token to only a few of them. The mechanism is fine-grained expert routing. Of 45 language layers, 42 are routed MoE layers, each holding 288 experts. For every token, a learned router selects the top 8 experts to run , that is 8 of 288 experts firing per layer, with the rest of the weights sitting idle for that token.
The model also uses a hybrid attention architecture that interleaves Sliding Window Attention (512-token window) and Global Attention at a 3:1 ratio, and uses Multi-Layer MTP (Multi-Token Prediction with 3 heads) for low-latency reasoning chains. MTP is a training and inference technique where the model learns to predict several future tokens simultaneously rather than one at a time , this enables a form of built-in speculative decoding that accelerates generation without needing a separate draft model.