Step 5 Preview
Step 5 Preview is StepFun’s 600B-total/27B-active mixture-of-experts model. Official downloadable FP8 and BF16 repositories now provide deployment material. The model card advertises up to 1M combined context; hosted access, self-hosted runtime limits and license conditions must be checked separately.
Inspect Step 5 checkpoint and API notes ↗Why it matters
A downloadable checkpoint makes local evaluation possible, while its memory requirements and conditional license determine whether that route fits a real deployment.
A weight milestone, not a second launch
The model’s September 20 introduction predates the October window. The official BF16 repository now contains weight shards and was updated on October 15. This is a downloadable-weight milestone, rather than evidence that the original model launched again or that every hosted account gained access.
FP8 and BF16 need different hardware
The card gives FP8 examples on four B300 GPUs or eight H200 GPUs. BF16 examples use eight B300 GPUs or sixteen H200 GPUs across two nodes. Weight memory, KV cache and runtime buffers are distinct. Selecting bfloat16 does not convert an FP8 checkpoint into BF16 weights.
- Deployment card
Read the exact numerical format and multi-node instructions.
Understand the commercial conditions
StepFun Model License 1.0 allows use and modification subject to conditions. Redistribution retains notices and the license. A product exceeding 100M active users or $20M revenue in a calendar month triggers attribution. Hardware makers and qualifying MaaS businesses can require separate commercial agreements; the MaaS revenue test covers aggregate affiliate revenue over twelve consecutive months. The license also specifies internal-use and certified-distribution exceptions.
- Governing license
Read the exact attribution, commercial-license and exception clauses.
Treat the reference server as an unexecuted recipe
The card’s vLLM/SGLang examples expose an OpenAI-compatible API and use step5 as the served request name. That local name is not evidence of a public StepFun API alias. The BF16 card explicitly says the complete standalone conversion-to-generation path has not been verified end to end.
- Model contract
Checkpoint IDs and documented request boundaries.
Keep evaluation settings visible
The publisher’s benchmark notes use High reasoning and named harnesses. They do not prove general financial, coding or visual accuracy. Keep task evidence, exact checkpoint, prompt and runtime settings together when evaluating your own workflow.
Useful first jobs
- A weight milestone, not a second launch
- FP8 and BF16 need different hardware
- Understand the commercial conditions
Frequently asked questions
Are the weights Apache-2.0?
No. The official repository uses StepFun Model License 1.0 with attribution and conditional commercial-license terms.
Does a 1M context claim guarantee a 1M-token local request?
No. Model context, KV-cache capacity, concurrency and runtime buffers must all fit the selected hardware.
Sources and evidence
Checked October 21, 2026 against the linked primary sources. Specifications and examples are publisher documentation; no independent model or hardware run is claimed.
- BF16 model card and files ↗
StepFun · Primary source
- FP8 model card ↗
StepFun · Primary source
- StepFun Model License 1.0 ↗
StepFun · Primary source