# step-5-preview decision evidence

Step 5 Preview is StepFun's sparse mixture-of-experts flagship (600B total, 27B active per token) with a 1M-token context window and text, image and video input. It is API-only today; StepFun has announced open weights for 15 October, and the reviewed tests are mixed rather than one-sided.

Verified: 2026-09-24; snapshot: 2026-09-25.

## Published model card

As the vendor publishes it for the model and its weights; these are not the limits of a hosted endpoint.

- Context window: 1M tokens
- Maximum output: 64k tokens
- Modalities: Text, image and video input; text output
- Parameter count: 600B total, 27B active per token (sparse MoE)

## Open weights and deployment

Status: Unavailable. StepFun had not published weights as of 2026-09-24. Its announcement says the model will be released with open weights on October 15, and the ModelScope pre-release page lists 2026-10-15; the licence and recommended runtime are not yet published. Step-5-Preview repositories under other accounts are not official.

Open-weight alternatives: [deepseek-flash](https://dash.chinaapi.ai/models/deepseek-flash/deploy), [glm-5.3-flash](https://dash.chinaapi.ai/models/glm-5.3-flash/deploy)

## Real-task reports

### [Step 5 Preview and GPT-5.6 Sol on a Blender-to-3D-print task](https://x.com/MinLiBuilds/status/2101614012769439966)

By [实践哥 Li](https://x.com/MinLiBuilds) (@MinLiBuilds). X creator sharing hands-on agent tests who took part in StepFun's early evaluation of Step 5 Preview

Published: 2026-09-20; retrieved: 2026-09-24; verified: 2026-09-24.

- Task: Model an 8 cm spring caterpillar in Blender from a reference image, then 3D-print it
- Method: Both models ran in an identical environment (pi coding agent, Computer Use, Blender 5.2.1, 1M context, maximum thinking), were timed end to end, and their STL files were printed.
- Finding: Step 5 Preview took 54 min 40 s against GPT-5.6 Sol's 14 min 34 s but gave the model a flat base that printed cleanly; Sol's version stood on a few small feet and nearly failed with stringing.
- Limitation: The author took part in StepFun's early evaluation. In a second round that spelled out printability in the prompt, both models did well, so the gap appeared only when the prompt left it implicit.
- Step 5 Preview time: 54 min 40 s
- GPT-5.6 Sol time: 14 min 34 s

### [Step 5 Preview and Claude Fable 5.1 on one front-end prompt](https://x.com/notjazii/status/2101653125623152714)

By [J A Z I I](https://x.com/notjazii) (@notjazii). X account that tests AI models and covers AI news

Published: 2026-09-20; retrieved: 2026-09-24; verified: 2026-09-24.

- Task: Build a front-end page from a single prompt
- Method: Same prompt at the highest available reasoning setting for both models; the author reported time and cost and repeated the Step 5 run without skills.
- Finding: Step 5 Preview finished in 10 minutes for about $0.50; Claude Fable 5.1 took 40 minutes and $17. The repeated Step 5 run came out almost the same.
- Limitation: One prompt; the post leaves the quality comparison to readers rather than scoring it, and the cost gap follows each vendor's list prices.
- Step 5 Preview: 10 min · $0.50
- Claude Fable 5.1: 40 min · $17

### [Step 5 Preview and DeepSeek V4.1 Flash on a 3D arcade game](https://x.com/ItsmeAjayKV/status/2102749825469219220)

By [AJ](https://x.com/ItsmeAjayKV) (@ItsmeAjayKV). X user focused on local LLMs who names DeepSeek V4.1 Flash as their default model

Published: 2026-09-23; retrieved: 2026-09-24; verified: 2026-09-24.

- Task: Build Katamari Tiny, a 3D arcade game, from a single prompt
- Method: Same prompt through each vendor's API with high thinking effort; the author posted the outputs side by side and reported tokens used.
- Finding: Step 5 Preview used 39,830 tokens against DeepSeek V4.1 Flash's 26,898, and the author judged DeepSeek's game faster and better.
- Limitation: One prompt judged by an author who names DeepSeek V4.1 Flash as their default model; the post gives no timings or quality rubric.
- Tokens: 39,830

### [Step 5 Preview and DeepSeek V4.1 Flash on one Three.js scene](https://x.com/ItsmeAjayKV/status/2101598311706943836)

By [AJ](https://x.com/ItsmeAjayKV) (@ItsmeAjayKV). X user focused on local LLMs who names DeepSeek V4.1 Flash as their default model

Published: 2026-09-20; retrieved: 2026-09-24; verified: 2026-09-24.

- Task: A Three.js scene of giant red mesas in a desert, with long shadows and a huge empty sky
- Method: Same prompt at high reasoning for both models, one shot to a single HTML file with no iterations and no harness; the author compared the results by eye.
- Finding: Step 5 Preview produced the richer scene, with better atmosphere, a tracked sun, more particle and dust effects and on-screen text; DeepSeek V4.1 Flash's version was more restrained but had clearly better ground and mesa textures and a smoother camera.
- Limitation: One prompt judged by eye by the author; the post does not say how either model was accessed.
- Attempts: One shot each

### [Step 5 Preview on the pelican-bicycle SVG](https://x.com/Fei2411/status/2101641124440084552)

By [FeiZ](https://x.com/Fei2411) (@Fei2411). Agent developer on X who lists himself as a Cognition ambassador

Published: 2026-09-20; retrieved: 2026-09-24; verified: 2026-09-24.

- Task: An HTML page with a 2D SVG animation of a pelican riding a bicycle
- Method: Run in OpenCode V2; a disconnection meant a second prompt asking for more detail, so it was not a strict one-shot.
- Finding: The scene looked good overall, roughly on a par with Kimi K3 in the author's view.
- Limitation: One prompt with a follow-up after a disconnection, judged by eye.
- Prompts: 2

### [Step 5 Preview on kzhu's six-question test](https://www.youtube.com/watch?v=jNnIwNo4YwM)

By [AI产品狙击手 kzhu](https://www.youtube.com/@kevinzhu9305) (@kevinzhu9305). Chinese YouTube reviewer who scores models on a six-question test and publishes every score with per-question notes

Published: 2026-09-20; retrieved: 2026-09-24; verified: 2026-09-24.

- Task: The author's six-question test: constrained sentences, pelican and compound-bow SVGs, a browser OS with a space game, a 3D racing game and a Trello board
- Method: Step 5 Preview run in Pi at maximum thinking; the finding is the author's own summary in the video description.
- Finding: Most sentences were fluent, the pelican and compound-bow SVGs were strong with only small detail issues, the browser OS was complete and its space game playable, the 3D racing game was refined, and the Trello board's interactions were smooth apart from one small UI flaw; the author rated the overall performance outstanding.
- Limitation: A summary without scores; Step 5 Preview does not appear on the author's leaderboard, so it cannot be compared with the scored rows.
- Harness: Pi, maximum thinking

## Public leaderboard coverage

A board missing from this list has no figure for this model in this snapshot; missing coverage is not a zero.

- Artificial Analysis Intelligence Index — Intelligence Index: 44 (retrieved 2026-09-21) [source](https://artificialanalysis.ai/leaderboards/models)
- Artificial Analysis · Output Speed — Median output tokens/s: 83 (retrieved 2026-09-22) [source](https://artificialanalysis.ai/leaderboards/models)
- Artificial Analysis · GDPval-AA v2.1 — Agentic Real-World Work Tasks, (Elo-500)/2000 (%): 53 (retrieved 2026-09-21) [source](https://artificialanalysis.ai/leaderboards/models)
- Artificial Analysis · Terminal-Bench v4.0 — Agentic Coding & Terminal Use (%): 33 (retrieved 2026-09-21) [source](https://artificialanalysis.ai/leaderboards/models)
- Artificial Analysis · MLCR-AA — Medical Long Context Reasoning (%): 16.7 (retrieved 2026-09-22) [source](https://artificialanalysis.ai/evaluations/mlcr-aa)

## Curated comparisons

- [claude-fable-5-1 vs step-5-preview](https://dash.chinaapi.ai/models/compare/claude-fable-5-1-vs-step-5-preview)
- [deepseek-flash vs step-5-preview](https://dash.chinaapi.ai/models/compare/deepseek-flash-vs-step-5-preview)

## Official sources

- [Step 5 Preview: Advancing the Pareto Frontier](https://www.stepfun.com/step-5-preview) — StepFun, retrieved 2026-09-24
- [Step 5 Preview model documentation](https://platform.stepfun.ai/docs/en/guides/models/step-5-preview) — StepFun, retrieved 2026-09-24