Chinese Step 5 Preview can be downloaded and run on your own hardware

Step 5 Preview is StepFun's flagship model for what is known as agentic work. That means the AI solves tasks on its own in multiple steps, such as writing and fixing code. According to the model card, its strengths are software development, long-running tasks, professional knowledge work and finance.

The model is a so-called sparse Mixture-of-Experts model. It has 600 billion parameters in total, but only 27 billion are active per token (a token is a small chunk of text). That makes it easier to run than the total figure suggests. However, vLLM lists 604 billion parameters, so the sources do not fully agree. According to the model card, the model can handle up to 1 million tokens of context, meaning very long texts at a time, and accepts both text and images as input. Some search results also mention video, but that is not stated in the model card.

Open weights and license

The weights, meaning the trained model itself, are available on Hugging Face and ModelScope in both FP8 and BF16, two different number formats. The license is StepFun Model License 1.0. The model card describes it as similar to Apache 2.0, but with additional attribution requirements and conditions for commercial use. Developers Digest wrote about the license before the weights were released and called it unconfirmed at the time.

According to Developers Digest, the model first arrived as an API in September 2026 and on OpenRouter on October 8. StepFun had promised open weights on October 15, 2026. The actual date of the weight release is not stated directly in the model card.

Requires powerful hardware

The model can be run with vLLM, SGLang and Transformers. According to vLLM, a single node requires at least 4 NVIDIA B300 GPUs, or 8 H200s for FP8. There is also an NVFP4 variant, where the maximum context drops to 320,000 tokens. This is equipment for companies and research environments, not for ordinary computers.

On OpenRouter, the model costs USD 1.00 per million input tokens and USD 2.70 per million output tokens.

How does it perform?

The Artificial Analysis Intelligence Index gives the model a score of 44 and a rank of 40th out of 226. It used about 160 million output tokens in the test and is described as very verbose. Early telemetry from OpenRouter showed roughly 17% errors in structured output.

Vendor-reported benchmarks show 67.7% on DeepSWE v1.1 versus 74.0% for Claude Opus 5, and 80.5% on ProgramBench versus 82.3%. The figures come from the vendor or third-party compilations and are not verified. According to Developers Digest, there is still a noticeable gap to the best models on long-running tasks.