StepFun is making its latest model announcement around a practical question: how much work can an AI system carry through before it loses the thread? Its new Step 5 Preview is designed for agentic tasks, software engineering, professional knowledge work, and finance, according to the company. The model’s headline specifications are large, but its architecture is designed to avoid activating the full parameter count for every token.
On its official announcement page, StepFun describes Step 5 Preview as a sparse Mixture-of-Experts model with 600 billion total parameters and 27 billion active parameters per token. The company says it supports a one-million-token context window and vision input. Its documentation lists text, image, and video input with text output, while the model is available through StepFun products and its API.
The technical claims are largely StepFun’s own, although the independent evaluator Artificial Analysis lists the model with 600 billion parameters, a one-million-token context, text and image inputs, and an Intelligence Index score of 44. The key distinction is that 600 billion is the total parameter count, not the number the system activates for every token. The 27-billion figure is the active-per-token count that StepFun provides for its sparse architecture.
A sparse architecture targets long-running AI work
A Mixture-of-Experts design routes each token through only part of a larger model. StepFun is using that architecture to argue that Step 5 Preview can combine a large overall model with a more selective active path. The company’s stated goal is not simply a high benchmark score. It is a model capable of sustained work across coding, research, financial analysis, and tool-using agent tasks.
The one-million-token context window is central to that positioning. In simple terms, a larger context can allow a system to consider more material in one interaction, such as long code repositories, extensive technical documentation, or a large research corpus. Context length alone does not guarantee that a model will reason correctly across all that material. It does, however, set a ceiling on the amount of input the model can accept in a single workflow.
StepFun’s documentation gives a more granular picture than the launch page. It lists a maximum one-million-token input and a maximum 64,000-token output. The figures should not be combined into an unsupported claim about a single 1.064-million-token context. They describe separate documented input and output limits.
The company also says the model accepts text, images, and video, while its main launch announcement specifically mentions vision input. Video support is an official StepFun documentation claim, not independently corroborated in the same way as text and image inputs. That difference is worth preserving because model capabilities can vary across interfaces, versions, and product settings.
The release follows a growing Chinese focus on models that can complete multi-stage tasks rather than only answer questions. EastFrontier recently examined Xiaomi’s window into MiMo V2.6 reinforcement-learning training, another example of how Chinese AI firms are increasingly discussing the training and execution methods behind their systems. StepFun’s emphasis is on what happens after inference begins: keeping track of actions, feedback, and revisions over a longer horizon.
Benchmark tables need to stay in their proper context
StepFun presents a broad table of results across coding, agent, reasoning, finance, and multimodal tasks. Some figures are drawn from public benchmarks, while others involve benchmarks or case studies developed or described by the company. They are useful evidence of the performance StepFun is claiming, but they are not a substitute for independent replication.
The company reports a 67.7% result on DeepSWE v1.1, 49.0% on its StepCodeBench evaluation, 80.5% on ProgramBench, 33.3% on Terminal-Bench v4, 29.5% on Agents’ Last Exam command-line tasks, 66.4 on FrontierFinance, and 83.3% on DRACO. StepFun says its StepCodeBench covers 553 repositories, nine task categories, 20 application domains, and 33 programming languages. Because that benchmark is internally developed, the result should be read as a company evaluation rather than an independent ranking.
The comparison table itself also calls for care. StepFun’s 33.3% Terminal-Bench v4 score sits below several named competitors in the displayed table. On another result, HLE with tools, the company warns that Step 5 Preview and one comparator were tested on a text-only subset while other models were assessed on the full data set. Those settings are not directly comparable.
The independent Artificial Analysis score of 44 provides a narrower external reference point. It does not independently validate every result in StepFun’s much larger comparison table. Nor does it establish the company’s broader claims that the model sits on a new cost-performance frontier. The appropriate phrasing is that StepFun reports its results, while Artificial Analysis independently lists a score of 44 on its own index.
StepFun also describes long-horizon demonstrations. In one, it says the model spent roughly 22 hours optimizing an MLA GPU kernel on an Nvidia H100 and reached a reported peak of 508 TFLOPS. In another, it says an automated post-training workflow improved a Qwen3-30B-A3B model’s AIME24 performance from 53.3% to 60%. These are company demonstrations with defined setups, not independently reproduced general results.
The open-weight plan is still a future commitment
StepFun says Step 5 Preview is available through its products and API now and that the model will be released with open weights on October 15. That is a planned date, not evidence that weights are already public. Artificial Analysis listed the model as proprietary when it was checked for this research, and the sources reviewed do not establish licensing, commercial-use terms, redistribution rights, fine-tuning permissions, or self-hosting requirements.
That difference matters. “Open weights” and “open source” are not interchangeable descriptions unless a license and distribution terms are actually available. The October 15 commitment may be important if StepFun follows through, especially for developers who want to inspect or adapt a large Chinese model. For now, the model should be described as a hosted offering with a stated future open-weight plan.
The exact public launch date is also less tidy than the technical announcement. StepFun’s primary page does not display a calendar date. A separate independent report places API availability on September 20, while Artificial Analysis lists a September release and gives September 18 in its FAQ. The sources do not reconcile those differences. The safest formulation is that Step 5 Preview was publicly available through StepFun’s hosted products and API by the September 21 research date, with September 20 independently reported for API access.
StepFun’s move belongs in the wider competition among Chinese firms seeking to pair large models with practical agent behavior. EastFrontier’s coverage of Moonshot connecting Kimi to financial data for institutional users showed a different approach: combining a model with specialized financial information. StepFun is instead stressing model architecture, long context, and repeatable execution across a wider range of work.
The announcement does not prove that Step 5 Preview has solved the hardest reliability problems in agentic AI. It does make the company’s strategy clear. StepFun is betting that developers will value a model that can keep working through long, tool-rich tasks while activating only a fraction of its total parameters at each step. The remaining questions are empirical: how the model performs outside company evaluations, how its API is adopted, and whether the October open-weight release arrives with terms that developers can use.