StepFun launches Step 5 Preview, a 600B-parameter sparse model with a million-token context, and promises open weights on October 15

StepFun has released Step 5 Preview, its new flagship model for agentic work, targeting software engineering, professional knowledge work and finance. The company's main pitch is cost. It states that the model delivers comparable intelligence at a substantially lower cost per task, which is the Pareto frontier framing used in the launch title. The model is available now as a hosted API and on StepFun's own platform. Self-hosting must wait for open weights, which StepFun says will land on October 15, 2026. Step 5 Preview is a sparse mixture-of-experts model with about 600 billion total parameters, of which roughly 27 billion are active per token, or about 4.5 percent of the weights. MarkTechPost notes that 600 billion parameters need about 1.2 terabytes in BF16 before accounting for the KV cache, so organisations planning to self-host should expect multi-GPU server hardware once weights ship. The official documentation lists the model ID as step-5-preview with a context window of one million tokens. It accepts text, images and video as input and produces text output. Reasoning effort can be set to low, medium or high, and the API supports streaming, tool calling, JSON mode, JSON schema and prompt caching. On research tasks, StepFun says the model coordinated 950 web fetches in a single run, though the excerpt does not give the full result. The release adds another large Chinese open-weight candidate to a field that already includes Qwen and others, with the weights themselves still weeks away.