Shanghai-based StepFun has released the complete model weights of Step 5 Preview as open source, making good on the pledge it made when the model launched on September 20. The release, reported by Chinese tech media including CNMO on October 15, turns one of the most closely watched Chinese frontier models into a fully downloadable artifact alongside its hosted API, which has been open to third-party integration since launch.
The architecture is a sparse mixture-of-experts: 600 billion total parameters with only 27 billion active per token, a ratio that keeps per-token compute close to a small dense model while retaining the capacity of a frontier system. The model supports a 1 million-token context window and natively accepts both text and image input, with tool calling and configurable reasoning depth aimed squarely at long-running agentic workloads such as software engineering, professional knowledge work and financial analysis.
On the benchmark side, Step 5 Preview scores 44 on the Artificial Analysis Intelligence Index, which the company says places it among the top three open-weight models globally, behind only Qwen3.8 Max (0902) and GLM-5.3 (max). Artificial Analysis figures cited by the company put the cost of completing one index task at about $0.71, versus $5.86 for Claude Opus 5 — the basis for StepFun's claim that matching tasks cost roughly one-eighth as much as Anthropic's flagship. Those are company-cited numbers, not independent audits, and the benchmark table is self-reported.
Pricing on the hosted API is $1.00 per million input tokens and $2.70 per million output tokens, with cache reads around $0.05 — well below the frontier average on both axes. OpenRouter currently serves as the main third-party gateway for the model, and its weighted average input price has been sitting far below list thanks to aggressive cache-hit behavior on agentic workloads.
The open-weights release lands mid-way through a two-week promotion that runs October 12 to 25, during which Step 5 Preview is available free and exclusively inside Tencent's WorkBuddy office agent. The sequencing is deliberate: enterprise users get the hosted model through a distribution partner while developers can download and self-host the same system, a two-track approach that mirrors what DeepSeek and Qwen have used to build developer gravity.
The strategic read is straightforward. StepFun is positioning efficiency — not raw scale — as the next scaling law, arguing that the era of buying intelligence with ever more compute is giving way to one where cost per completed task is the metric that matters. Whether the open weights hold their published performance outside the vendor's own evaluation harness will be the real test, but the release gives independent labs everything they need to find out.
Comments (0)
Log in to join the discussion
Log InNo comments yet