What is Step 5 Preview?
Step 5 Preview is StepFun's flagship model, launched on its API on 20 September 2026 with weights following on Hugging Face and ModelScope in mid-October. It is a sparse mixture-of-experts model with 600 billion parameters, 27 billion active per token, and a context of up to 1 million tokens. StepFun says it reads text, images and video and writes text, and pitches it at long-running agent work, software engineering and finance.
The weights come in BF16, FP8 and NVFP4 under the StepFun Model License 1.0, which StepFun describes as Apache-2.0-style with extra conditions. Products with over 100 million monthly users or $20 million in monthly revenue must display "Powered by StepFun AI", and device makers and model-hosting businesses with over $20 million in yearly revenue need a separate commercial licence. Internal use is exempt.
On StepFun's API it costs $1.00 per million input tokens and $2.70 per million output tokens, with a 90% discount on cached input. OpenRouter and DeepInfra also serve it at the same price.
How much does Step 5 Preview cost?
| Plan | Price | What you get |
|---|---|---|
| Weights | Free | StepFun Model License 1.0 |
| StepFun API | $1.00 in / $2.70 out per 1M tokens | 90% off cached input |
Open weights under the StepFun Model License 1.0. StepFun API: $1.00 input / $2.70 output per 1M tokens, 90% off cached input; same price on OpenRouter and DeepInfra.
Is Step 5 Preview worth it?
Step 5 Preview is good value rather than a leader. Artificial Analysis scores it 44 on its Intelligence Index, level with Kimi K3 and a point behind GLM-5.3, at $1.03 per index task against Kimi K3's $2.00. It is strong on knowledge and long documents: 46% on Humanity's Last Exam, in line with Kimi K3, and 88.3% on Artificial Analysis's long-context reasoning test.
The weakness is the work StepFun markets it for. On Artificial Analysis's agent tests it trails GLM-5.3 and Qwen3.8-Max: 33% on Terminal-Bench 4.0 against 42% and 39%, and lower scores on the GDPval and Briefcase office-task tests. It also gave wrong answers more often than GLM-5.3 on the Omniscience test, 43% against 30%. Xiaomi's MiMo-V2.6-Pro, released a day later, scores two points higher for less money.
Its practical advantage is size. With 27 billion active parameters from 600 billion, the FP8 checkpoint runs on a single node of four B300 or eight H200 GPUs, where MiMo-V2.6-Pro and Kimi K3 need far more. For a team that wants a near-frontier open model on one server, that matters more than two index points.
Read the licence and data terms before committing. The display and commercial-licence conditions only apply to large products, device makers and model-hosting businesses. StepFun's API terms let it train on prompts from its Startup and early-access programmes, and say nothing explicit about ordinary paid traffic.
Where it earns its place
A financial research team
Analyse long filings, charts and spreadsheets in one pass
StepFun pitches finance as its strongest area, and the 1M-token context and 88.3% long-context reasoning score suit whole-filing analysis. At $1.00/$2.70, a 400K-token document with a 10K-token report costs about $0.43.
A company that wants a frontier-class open model on one server
Self-host without a multi-node cluster
Serve the FP8 checkpoint with vLLM or SGLang on one node of four B300 or eight H200 GPUs. Internal use is outside the licence's commercial conditions.
A developer comparing Chinese open models on price
Get Kimi K3-level quality for less
Step 5 Preview matches Kimi K3 on Artificial Analysis at about half the cost per task. For agent-heavy work, GLM-5.3 or MiMo-V2.6-Pro score higher.
Before you commit
- Setup
- Call step-5-preview on StepFun's platform, OpenRouter or DeepInfra, or download the BF16, FP8 or NVFP4 weights; the FP8 checkpoint is gated. Reasoning effort is low, medium or high.
- Integrations
- OpenAI-compatible API; vLLM, SGLang and Transformers recipes on the model card; NVIDIA's NVFP4 checkpoint for Blackwell GPUs.
- Data and privacy
- Weights under the StepFun Model License 1.0: display rule above 100M users or $20M monthly revenue, commercial licence for device makers and MaaS businesses over $20M a year, internal use exempt. StepFun's API terms allow training on Startup and early-access programme traffic; terms for other paid traffic are not explicit.
- Cost at scale
- $1.00/$2.70 per million tokens with 90% off cached input. 1,000 agent tasks of 500K input and 30K output tokens cost about $581 uncached. Self-hosted FP8 needs four B300 or eight H200 GPUs; BF16 needs eight B300 or sixteen H200.
- Limits
- Up to 1M tokens of context. Artificial Analysis lists a $1.03 cost per index task; measured speeds vary from 92 to 229 tokens a second by provider. Still labelled a preview.
What's good
- Scores 44 on Artificial Analysis, level with Kimi K3, at about half the cost per task
- FP8 runs on one node of four B300 or eight H200 GPUs
- 1M-token context with strong long-document results
- Same $1.00/$2.70 price on StepFun, OpenRouter and DeepInfra
What's not
- Trails GLM-5.3 and Qwen3.8-Max on agent tests
- Higher wrong-answer rate than GLM-5.3 on Artificial Analysis's Omniscience test
- Licence needs a commercial agreement for device makers and large model hosts
- No technical report published yet
Best for
Teams that want a near-frontier open-weight model for long documents and research that fits on a single GPU server.
Step 5 Preview alternatives
Step 5 Preview: common questions
- Is Step 5 Preview worth it?
- Kimi K3's score at half its cost per task and four GPUs to self-host, but it slips on the agent tests StepFun itself leads with. We rate it 4.2 out of 5.
- How much does Step 5 Preview cost?
- Open weights under the StepFun Model License 1.0. StepFun API: $1.00 input / $2.70 output per 1M tokens, 90% off cached input; same price on OpenRouter and DeepInfra.
- What are the downsides of Step 5 Preview?
- Trails GLM-5.3 and Qwen3.8-Max on agent tests; higher wrong-answer rate than GLM-5.3 on Artificial Analysis's Omniscience test; licence needs a commercial agreement for device makers and large model hosts; no technical report published yet.
- Who is Step 5 Preview best for?
- Teams that want a near-frontier open-weight model for long documents and research that fits on a single GPU server.
- How much does Step 5 Preview cost for a team or in production?
- $1.00/$2.70 per million tokens with 90% off cached input. 1,000 agent tasks of 500K input and 30K output tokens cost about $581 uncached. Self-hosted FP8 needs four B300 or eight H200 GPUs; BF16 needs eight B300 or sixteen H200.
- Is Step 5 Preview safe for company data?
- Weights under the StepFun Model License 1.0: display rule above 100M users or $20M monthly revenue, commercial licence for device makers and MaaS businesses over $20M a year, internal use exempt. StepFun's API terms allow training on Startup and early-access programme traffic; terms for other paid traffic are not explicit.
- What are the best Step 5 Preview alternatives?
- The closest alternatives we have reviewed are MiMo-V2.6-Pro, GLM-5.3, Kimi K3 and Qwen3.8-Max.
Answers checked Oct 25, 2026.
Make Step 5 Preview? Show this rating on your site
