StepFun Launches Step 5 Preview: 600B-Parameter MoE Built for Long Agent Tasks

Flagship MoE, Scaled to 600B

On October 8, StepFun’s Step 5 Preview landed on OpenRouter. It’s StepFun’s new workhorse: a sparse MoE with 600B total parameters, only 27B active per token.

That’s over 3x the size of its two previous 196B-class Flash models. The company positions it as an agent flagship — for software engineering and knowledge work, especially finance.

1M Context, Restrained Pricing

Specs: 1M-token context window, up to 64K output tokens per call. Input supports text, images, and video; output is text. The API supports function calling and structured output.

Pricing per OpenRouter’s listing: $1 per million input tokens, $0.05 on cache hits, $2.70 per million output tokens.

Third-party stats show agents pounced immediately: long-horizon agents like OpenClaw, Hermes Agent, and Cline are already among its top callers. OpenRouter stats show P50 latency of 1.68 seconds and median throughput of 137 tokens/second.

Company Claims vs Independent Scores

StepFun self-reported two heavyweight test scores: 66.4 on FrontierFinance, 67.7% on DeepSWE v1.1. Third-party Artificial Analysis gave it 44 on its intelligence index.

Those numbers come from different mouths. Independent evals haven’t caught up — reserve judgment until more tests land.

Open Weights October 15

Per reports, StepFun set October 15 for the Step 5 open-weight release. Its two previous Flash models (Step 3.5 Flash, Step 3.7 Flash) were open-sourced under Apache 2.0 — once weights drop, teams can self-host instead of paying per token.

StepFun was founded in April 2023, headquartered in Shanghai, one of China’s “AI six tigers.” By parameter scale, this one aims straight at long-horizon agent scenarios: read in an entire codebase or hundreds of pages of research, work through it with tools step by step, hand back the result. When the open weights land, we’ll see how deployable this MoE really is.