StepFun's Step 5 Preview weights hit Hugging Face early

StepFun's new flagship model is not supposed to be fully open yet. Step 5 Preview was announced on 20 September with a promise that the complete weights would follow on 15 October, but a public repository named Step-5-Preview-BF16 appeared on the company's Hugging Face account at 03:52 UTC the same morning, and it is not a placeholder. It holds 24 bfloat16 weight shards plus multi-token-prediction and vision-tower files, a weight index, a tokenizer and a chat template.

That is 28 safetensors files and roughly 1.2TB of checkpoints, sitting behind a download button anyone can press. The repository is listed as public and not gated, and the file listing resolves: the index, tokenizer and configuration all return HTTP 200.

What StepFun announced

According to ITHome, StepFun launched Step 5 Preview as its flagship foundation model aimed at coding, software engineering, professional knowledge work and finance, with the emphasis on agentic tasks that actually get finished. The company describes a sparse mixture-of-experts design with 600 billion total parameters and 27 billion active per token, a one-million-token context window, and native text and image input.

On the Artificial Analysis intelligence index the model is reported at 44, which puts it in the top three open models globally, at a reported per-task cost of about one-eighth of Anthropic's Claude Opus 5. StepFun also says it built StepCodeBench for the evaluation, covering 553 separate code repositories, nine task types, 20 application domains and 33 programming languages.

What the published configuration shows

The configuration in the repository matches the headline numbers closely enough to be worth reading. It describes 92 layers running grouped-query attention with a 192-dimension head, 352 experts with eight active per token across 88 mixture-of-experts layers, a maximum position count of 1,048,576, a vision tower working on 728-pixel images, three multi-token-prediction layers and bfloat16 throughout. The weight index lists 2,449 tensors and just over 1.2 trillion bytes, which is the size you would expect from a 600-billion-parameter model stored at full precision.

No model card, no licence

What the repository does not have is a README or any licence metadata. The model card URL returns a 404 and no licence identifier is published, so a team that downloads the checkpoint has no stated terms for using it. There are also no benchmark tables from StepFun in the repository itself, which means the numbers doing the rounds come from the launch material and the leaderboard rather than from the weights.

Why the two dates matter

Open-weight releases usually run the other way round. A lab publishes an API-only preview, then drops the weights weeks later once the hosted version has been monetised. StepFun kept that sequence in its announcement and then broke it within hours, which is good news for anyone who wants to run the model locally and a small headache for anyone who wants to know what they are allowed to do with it afterwards.

Our opinion

Publishing 1.2TB of frontier-adjacent weights with no licence is the open-weights equivalent of leaving the keys in the car and the paperwork at home. StepFun gets the goodwill of an early release and the leaderboard coverage that comes with it, while every commercial team that wants to build on Step 5 has to guess whether the weights are Apache, a bespoke research licence or something that only permits evaluation. That guessing game is not free. Legal departments do not sign off on maybes, so the missing licence quietly narrows the audience for an otherwise impressive drop.

It is still a better problem than the closed alternative, and the timing is sharp: the same model family lands within days of reports that China's open-weight labs are setting the pace on cost per task. But a licence file is a one-page document. StepFun had the time to write one, and the omission is the first thing a serious adopter will notice.