Overview
Step 5 Preview is StepFun's flagship model for agent work, released on 18 September 2026 with 600 billion parameters. It reads text, images and video and writes text back, and StepFun aims it at software engineering, long-document work and professional knowledge tasks. The model is served through StepFun's developer platform under the API id step-5-preview, and its weights are published on Hugging Face as stepfun-ai/Step-5-Preview (FP8) and Step-5-Preview-BF16, with an NVFP4 build added on 14 October 2026.
Step 5 Preview is a sparse mixture-of-experts model with 600B total and 27B active parameters. The weights ship under the StepFun Model License 1.0, which the model card describes as an Apache-2.0-style license with additional attribution and commercial-license conditions; the download is not gated, though the card notes that some checkpoints may need a Hugging Face login.
The context limit is 1M tokens on both the input and the output side, so a large codebase or a stack of documents can go into a single call without being split up first. Reasoning effort is selectable per request at low, medium or high, on the same model id, which lets a caller trade cost against depth without switching models.
Artificial Analysis, which benchmarks models independently, scores Step 5 Preview at 44 on its Intelligence Index — 24th of the 200 models it tracks, against a median of 24 — and measures output at 99.8 tokens per second with a 2.96 second time to first token. It also records the model as very verbose, generating 160 million tokens across its evaluation set against a median of 92 million.
| Released | 2026-09-18 |
|---|---|
| License | StepFun Model License 1.0 |
| Weights | Open weights |
| Parameters | 600B total / 27B active (MoE) |
| Context | 1M |
| Max output | 1M |
| Modalities | Text, Vision, Video |
| Status | Preview |
Benchmarks
Scores on a 0–100 scale (25-point gridlines); higher is better. Each benchmark links to its published source.
Pricing
| Input | $1.00 / 1M tokens |
|---|---|
| Output | $2.70 / 1M tokens |
Strengths
- 1M-token limit applies to output as well as input, not just the prompt
- Accepts video as well as images, with up to 60 images per request
- Low, medium and high reasoning effort selectable on the same model id
- $1.00 per million input tokens for an Artificial Analysis Intelligence Index score of 44
Best for
- Coding agents that read a whole repository in one call
- Long-document analysis and research over large PDF sets
- Screenshot, chart and video understanding inside an agent loop
- Financial and professional knowledge work, the areas StepFun targets
How to access
| Provider | Model ID |
|---|---|
| StepFun API ↗ | step-5-preview |
FAQ
Is Step 5 Preview open weights?
Yes. StepFun publishes the weights on Hugging Face as stepfun-ai/Step-5-Preview (FP8) and stepfun-ai/Step-5-Preview-BF16, with an NVFP4 build added on 14 October 2026, under the StepFun Model License 1.0 — an Apache-2.0-style license with additional attribution and commercial-license conditions, so it is open weights rather than a standard open-source license. The same model is also served through StepFun's API.
What image and video formats does Step 5 Preview accept?
Step 5 Preview accepts JPG and JPEG, PNG, WebP and static GIF images, with a limit of 60 images per request. For video, StepFun's documentation lists MP4, QuickTime and Matroska. Output is text only, so Step 5 Preview can describe or reason about a screenshot, chart or clip but cannot generate one.
How much does Step 5 Preview cost?
Artificial Analysis lists Step 5 Preview at $1.00 per million input tokens and $2.70 per million output tokens. Because Artificial Analysis also records the model as verbose — 160 million tokens generated across its evaluation set against a median of 92 million — output cost matters more here than the headline input price suggests.
How fast is Step 5 Preview?
Artificial Analysis measured Step 5 Preview at 99.8 output tokens per second, which ranks 45th of the 200 models it tracks and sits above the 70.3 tokens per second median. Time to first token is 2.96 seconds, so a request spends roughly three seconds before the first token arrives.