StepFun's Step 3.7 Flash Runs Agentic Loops at One-Ninth the Cost of Claude
StepFun open-sources a 198B MoE vision-language model that runs like an 11B model, targets agentic coding loops at a fraction of frontier costs
- 198B MoE, 11B active: Step 3.7 Flash activates only ~11B parameters per token, delivering up to 400 tokens/sec at Flash-tier cost.
- First multimodal in the Flash line: Adds a 1.8B Vision Transformer for native image, chart, UI, and document understanding — no separate vision API needed.
- Advisor Mode cuts agentic costs 9x: StepFun reports $0.19/task vs $1.76 for Claude Opus 4.6 on SWE-Bench Verified, at 97% of its coding performance.
- Leads ClawEval-1.1 at 67.1%: Outperforms DeepSeek V4 Pro (59.8%) on adversarial multi-turn agent safety and tool orchestration.
- Open weights under Apache 2.0: Available in BF16, FP8, NVFP4, and GGUF; runs locally on 128GB unified memory (Mac Studio, DGX Spark).
- Priced at $0.20/$1.15 per M tokens (in/out): Available now on Hugging Face, OpenRouter, NVIDIA NIM, and StepFun's platform.
Step 3.7 Flash is StepFun's latest open-weight model, and it is built around a single premise: running agentic loops at scale should not cost a fortune. The model is a 198B-parameter sparse Mixture-of-Experts (MoE) vision-language model that activates only about 11B parameters per token during inference. That gap between total and active parameters is the entire cost story.
A 198B model that runs like an 11B one
In MoE architectures, only a subset of "expert" sub-networks fires per forward pass. This keeps inference compute closer to an 11B dense model while maintaining a 198B total parameter budget. The practical result: it delivers a throughput of up to 400 tokens per second.
Unlike traditional dense models, Step 3.5 Flash (the predecessor) uses a fine-grained routing strategy with 288 routed experts per layer plus one shared expert that is always active, selecting only the top-8 experts per token. The model retains the "memory" of a 196B parameter model but executes with the speed of an 11B model. Step 3.7 Flash carries this same architecture forward, adding a new critical piece.
Step 3.5 Flash was text-only. Step 3.7 Flash adds a dedicated 1.8B Vision Transformer that injects image representations into the language context, enabling it to parse charts, PDFs, UI wireframes, and app GUIs without a separate vision API call.
The real headline: agentic cost efficiency
Benchmarks are useful, but the number that matters most for production teams is cost per agentic task. 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 , for teams running thousands of agent loops a month, that ratio is the headline.
StepFun achieves this through a feature they call Advisor Mode. The model runs the agentic loop end-to-end , calling tools, reading results, iterating , and escalates to a larger advisor model only at specific inflection points, such as planning or recovering from repeated failures. Most of the run stays at executor cost. With Advisor Mode enabled on SWE-Bench Verified, StepFun reports Step 3.7 Flash reaches 97% of Claude Opus 4.6's coding performance at roughly one-ninth the per-task cost ($0.19 vs. $1.76 per task). These are StepFun's internal figures.
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
