Launched via API in September, weights released later
StepFun (阶跃星辰), an AI company based in Shanghai, China, announced on X on Oct. 18 (Oct. 19 Japan time) that it had released the weights (trained parameters) of its large model "Step 5 Preview," and said it would hold a launch event in San Francisco on Oct. 22. Step 5 Preview is a model the company released via API on Sept. 20, and it had said it would release the weights on Oct. 15. With the weights released, companies can run the model on their own servers.
The model is a "mixture of experts (MoE)" model that uses only some of its many expert components, with 600 billion total parameters, of which 27 billion are used to process each token. Its context window is 1 million tokens, with output of up to 64,000 tokens. In addition to text, it can read up to 60 images or videos at once. It is mainly aimed at AI agents that carry out long sequences of steps on their own.
StepFun was founded in Shanghai in April 2023 by engineers including former Microsoft employees. China's Tencent and a Shanghai government-affiliated fund are among its investors, and Bloomberg reported in February that the company is aiming to list on the Hong Kong Stock Exchange.
Slightly ahead of Chinese rivals, a gap with top U.S. models
In evaluations announced by StepFun, Step 5 Preview scored 67.7 on DeepSWE (v1.1), which solves software development tasks. That was slightly above the 67.5 of "Kimi K3" from China's Moonshot AI and the 66.9 of "GLM-5.3" from China's Zhipu. However, there is a gap with OpenAI's "GPT-6 Astra" at 74.1 and Anthropic's "Claude Opus 5" at 74.0.
Ahead on a finance evaluation, but all evaluations self-measured
On Terminal-Bench v4, which solves command-line tasks, it scored 33.3, below GLM-5.3's 41.9 and GPT-6 Astra's 57.9. On FrontierFinance, which measures financial work, it scored 66.4, above GPT-6 Astra's 55.0 but short of Claude Opus 5's 69.7. All figures were measured by StepFun itself, and it has not yet published a technical report. One of the evaluations, "StepCodeBench," is a metric the company created itself.
| Evaluation | Step 5 Preview | Kimi K3 | GLM-5.3 | GPT-6 Astra | Claude Opus 5 |
|---|---|---|---|---|---|
| DeepSWE v1.1 | 67.7 | 67.5 | 66.9 | 74.1 | 74.0 |
| StepCodeBench | 49.0 | 43.9 | 40.2 | 61.0 | 63.9 |
| ProgramBench | 80.5 | 77.8 | 72.0 | 85.4 | 82.3 |
| Terminal-Bench v4 | 33.3 | 12.6 | 41.9 | 57.9 | 52.3 |
| FrontierFinance | 66.4 | 62.6 | 64.1 | 55.0 | 69.7 |
API pricing of $1 for input and $2.70 for output
API pricing is $1.00 for input (cached input is $0.05) and $2.70 for output per million tokens, with the reasoning process counted as output. There is also a free tier with limits on the number of requests. According to eesel AI, users on the U.S. forum Hacker News have said answers are verbose and the thinking process is too long. Using 27 billion parameters per processing step, more than rivals' small models, is also a disadvantage in terms of running costs.
