Step 5 Preview, a 1M-context MoE from StepFun, shows up on OpenRouter
9 October 2026 · filed under 1eaf0f76ffce
Telescope log. StepFun, a Chinese AI developer, has listed a new model, Step 5 Preview, on OpenRouter, the model-routing marketplace. The listing describes Step 5 Preview as a mixture-of-experts (MoE) architecture with a context window of one million tokens.
The appearance was first noted by users on Hacker News, where the posting circulated under the title “Step 5 Preview, a 1M-context MoE from StepFun, shows up on OpenRouter.” No further technical documentation, benchmark data, or official StepFun announcement accompanies the listing in the material reviewed here.
A context window of one million tokens would permit the model to process very large inputs in a single pass, such as lengthy documents, codebases, or multi-document collections, without the need to split material into smaller chunks. Mixture-of-experts architectures route individual queries through a subset of specialized internal parameters rather than activating an entire network, a design intended to manage computational cost at scale.
The source material does not specify StepFun’s founding, prior releases, pricing on OpenRouter, or independent verification of the stated context length. No statements from StepFun or OpenRouter are quoted in the available reporting. The listing’s presence on OpenRouter indicates the model is accessible to developers through that platform as of the preview’s appearance, though the preview label suggests the release is not yet final.