On October 15, StepFun put the weights of its flagship Step 5 Preview model on Hugging Face, making good on the promise it made at the September 20 launch to open-source the model on October 15. The repository currently holds two main weight sets, FP8 and BF16, plus an NVFP4 repository created on October 14. All three were updated this morning.
What the model card says
Step 5 Preview is a sparse mixture-of-experts (MoE) model with roughly 600 billion total parameters (the weight-file metadata shows 604B) and about 27 billion parameters activated per token. The context window reaches up to 1 million tokens, and maximum output is also listed at 1 million tokens. It accepts text and image input and produces text output. The model card positions it in a single sentence:
"Step 5 Preview is StepFun's flagship model for agentic work."
The license is StepFun's own StepFun Model License 1.0. The company calls it "Apache-2.0 style," but it adds an attribution requirement and conditions for commercial use. Commercial use at a certain scale requires a separate application, and the model card summary does not state the threshold, so you will need to read the full license text.
Benchmark results appear in the model card only as charts, covering coding, agentic tasks, professional knowledge work and finance, with no item-by-item figures in the text. The card notes that the scores were measured at the High reasoning setting, and that the GDPval-AA result cites Artificial Analysis data as of September 20. On the third-party side, Artificial Analysis gives the model 44 points on its Intelligence Index, placing it among the top three open-weight models.
What it means for teams in China
The most practical piece of information in this release is the deployment table in the model card:
| Framework | Precision | Hardware |
|---|---|---|
| vLLM / SGLang | FP8 | 1 B300 server (4 GPUs) or 1 H200 server (8 GPUs) |
| vLLM | BF16 | 1 B300 server (8 GPUs) or 2 H200 servers (16 GPUs) |
| Transformers offline | FP8 | 1 H200 server (8 GPUs) |
For companies in China, B300 systems are essentially unobtainable, and H200 supply into the Chinese market has been a stop-and-start affair this year, so the realistic route is most likely FP8 on an 8-GPU H200 server. The deployment table in the model card lists only NVIDIA's H200 and B300. StepFun has not yet offered an official plan for domestic accelerators, including whether they can run the model and how much performance would be lost.
The model card also flags two common pitfalls. First, a 1-million-token request may not fit in the KV cache even on the configurations above. Second, adding the --dtype bfloat16 flag to the FP8 weights does not convert the quantized weights back to BF16; to run BF16 you must download the separate BF16 weights. The Transformers loading path for sharded BF16 weights is marked by the company as not yet fully verified.
A rough calculation: at FP8, 600 billion parameters need about 600 GB for the weights alone, and eight 141 GB H200 cards offer about 1.1 TB of memory in total, leaving the rest for the KV cache. For teams that want to run the full model in their own data centers, a server like this is the entry ticket.
Twenty-five days from launch to open weights
At the September 20 launch, StepFun led with cost: the company said a Step 5 Preview task costs about one-eighth as much as the same task on Claude Opus 5. The model was then made available first through StepFun's own products and API, and in early October it climbed to No. 1 globally on OpenRouter's Trending chart.
With the weights out, the question shifts from whether the API is any good to whether self-hosting pays off. How soon the community produces low-bit versions such as GGUF, which would let a single machine or a small cluster run the model, will determine how many small and mid-sized teams this release can reach. StepFun also published a technical blog post at the same time, and how much detail it gives on the scale of the training data and the reinforcement learning recipe is something the community will have to check item by item.
Sources: StepFun Hugging Face model card, CocoLoop, STAR Market Daily, Ifanr. Parameter counts, deployment hardware and license terms follow the model card; the intelligence index score follows Artificial Analysis methodology.