CoAI
News

StepFun Releases Step 5 Preview, a Flagship Model for Agents With 600 Billion Total Parameters, and Plans to Release the Weights on October 15

On September 20, 2026, Chinese AI company StepFun (Jieyue Xingchen) announced its new flagship model Step 5 Preview and made it available via its API and other channels the same day. It is designed for uses in which AI carries out programming, specialized knowledge work and financial analysis while using tools on its own, and it can handle long inputs of 1 million tokens. The company says it will release the model weights on October 15.

September 20, 20265 min read

On September 20, 2026, Chinese AI company StepFun (Jieyue Xingchen) announced its new flagship model Step 5 Preview. It is a model for "agent" uses, in which AI carries out work on its own while using tools, and the company says it excels in programming, specialized knowledge work and finance. Alongside the announcement, it became available through StepFun's API and the company's own services. The model weights are scheduled to be released on October 15.

What is StepFun?

StepFun is an AI startup founded in April 2023 and headquartered in Shanghai. Although it is still little known in Japan, in China it is known as one of the leading companies developing large language models (AI that understands and generates text) in-house. In addition to text models, it also works on the StepAudio series for speech recognition and synthesis and music generation.

StepFun has previously announced Step 3.5 Flash in February 2026 and Step 3.7 Flash in May, releasing the model weights on Hugging Face. Releasing weights (open weights) means distributing the data of a trained model so that anyone can download it and run it in their own environment. Step 5 Preview is positioned as the top model following these.

Model features

According to StepFun's announcement, Step 5 Preview uses an architecture called MoE (Mixture of Experts). It has 600 billion (600B) parameters in total, but only 27 billion (27B) of them are used per pass. Even though the model as a whole is large, each pass runs only some of its "expert" parts, so it can run fast and cheaply for its size.

According to the developer documentation, the main specifications are as follows.

  • Amount it can read at once (context): 1 million tokens
  • Single output: up to 64K (about 64,000) tokens
  • Input: text, images and video (up to 60 images at a time)
  • Output: text
  • Depth of thinking: selectable from three levels, low, medium and high
  • Supports tool calling, structured output and caching (a mechanism that reduces costs by reusing identical input)

StepFun cites uses such as software development, organizing long documents, research and analysis, and reading the content of images and videos. It also explains that tools such as search and program execution are provided by the app that incorporates the model, and the model itself does not directly access computers or external services.

Performance evaluation

According to StepFun, Step 5 Preview scored 44 on the overall index (Intelligence Index) of Artificial Analysis, a third-party evaluation site that compares AI models, placing it "in the world's top 3 as an open-source model." However, the weights are scheduled for release on October 15, and Artificial Analysis currently treats it as a model without published weights. StepFun claims that the cost per task is one-eighth that of Anthropic's Claude Opus 5.

In StepFun's own published evaluations, the company says the model ranked just behind OpenAI's GPT-6 Astra or Claude Opus 5 and outperformed other open models on the command-operation portion of Agents' Last Exam, a comprehensive test of AI agents, on FrontierFinance, which measures financial investment research, and on DRACO, which measures cross-disciplinary research.

On programming, the company says that in evaluations by internal and external experts, about 70% of participants judged that the model "can independently solve programming tasks of medium to high complexity." It also covers a wide range: the company explains that it can create web interfaces and data visualizations, build 3D assets using the 3D creation software Blender, and, given development documentation and user permission, even develop and debug (fix problems in) programmable devices.

StepFun also gave an example of long-running work. In an experiment using one NVIDIA H100 in which models automatically improved a GPU program over 24 hours (each model ran four times, and the best result was adopted), the company says processing speed reached 508 TFLOPS in about 22 hours, exceeding Claude Opus 5's 493 TFLOPS. On the other hand, it also explains that in its in-house programming evaluation, a gap with the top models remains on the most difficult and time-consuming tasks. All of these are StepFun's own evaluations.

Where it is available and pricing

Step 5 Preview is available through the API on the developer sites for China and for overseas users (platform.stepfun.com and platform.stepfun.ai), and in Studio, which you can try in a browser. It can also be used with the monthly Step Plan and called from development tools such as Claude Code.

Pricing for the overseas API is $1.00 per million tokens for input, $0.10 for cached input and $2.70 for output (as of October 12, 2026). The output price includes the model's thinking process. For China, the company announced that the price of cached input will rise from 0.35 yuan to 0.70 yuan per million tokens from October 14 Beijing time. Users who have used the pay-as-you-go API before then will continue to get the original price. The monthly Step Plan is not subject to the price increase.

The company says it will release the model weights on October 15. However, the type of license and where the weights will be released had not been announced as of October 12. There is also no particular guidance on Japanese language support or availability for Japan.

CoAI's take

Step 5 Preview is a model that came out in the order of being made available via API first and having its weights released a little later. It is large at 600 billion parameters in total, but only 27 billion are active per pass, so once the weights are released, it becomes an option for companies with large GPU servers to run in their own environments.

On the other hand, many of the evaluation figures were measured by StepFun itself. Two days after the announcement, Xiaomi's MiMo-V2.6-Pro scored 46 on the same Artificial Analysis index, and the competition at the top of open models is shifting at short intervals (Step 5 Preview will be compared in the same category once its weights are released). Whether to actually use it is best decided by trying it on your own work after also checking the license terms to be shown on October 15.

Sources

Back to news