Step 5 Preview
StepFun's 600B agent model with 27B active parameters, a 1M context and image input, under a custom license with hardware and hosting terms.
In short. Step 5 Preview is a very large open model from the Chinese company StepFun, built for coding agents, long multi-step tasks and professional work such as finance. It has 600 billion parameters but uses 27 billion for each token, reads up to a million tokens and can look at images. Running it takes a multi-GPU server, and the license has special terms for hardware makers and hosting businesses.
Summary
StepFun launched Step 5 Preview as an API model on 20 September 2026 and published the open weights on 15 October 2026. It is a sparse mixture of experts with 600B parameters (604.3B on Hugging Face) and 27B active per token, a context of up to 1M tokens and an output limit of up to 1M tokens. It takes text and images; the card's introduction also mentions video, while its specification table lists only text and image. StepFun calls it its flagship model for agentic work, strongest in software engineering, professional knowledge work and finance. The license is the StepFun Model License 1.0, a custom license with extra terms for hardware makers and model-as-a-service businesses.
In StepFun's own charts Step 5 Preview ran at its High reasoning setting while its rivals ran at Max, so the comparison is not like for like. It scores 67.7 on DeepSWE v1.1 (Kimi K3 67.5, GLM-5.3 66.9, Claude Opus 5 74.0, GPT-6 Astra 74.1), 49.0 on StepCodeBench (Kimi K3 43.9, GLM-5.3 40.2, Claude Opus 5 63.9), 33.3 on Terminal-Bench 4.0 (GLM-5.3 41.9, Kimi K3 12.6) and 66.4 on FrontierFinance, ahead of Kimi K3 at 62.6 and behind Claude Opus 5 at 69.7. Its GDPval-AA v2.1 score is 1566. StepFun also built three finance benchmarks of its own. Except for GDPval-AA, which StepFun takes from Artificial Analysis, the scores are StepFun's.
StepFun lists vLLM and SGLang recipes and a standalone Transformers implementation, with the High reasoning setting and the recommended settings of temperature 1.0, top-p 0.95 and top-k 0. The FP8 weights are about 650 GB and run on one 8-GPU H200 node or one 4-GPU B300 node. The BF16 weights are about 1.2 TB and need two H200 nodes or one 8-GPU B300 node, and an NVFP4 build is 394 GB. In the 30 days to 11 November 2026, Hugging Face counted about 8,100 downloads of the main FP8 repository and about 1,600 each of the BF16 and NVFP4 ones. You can use, modify and distribute the model commercially for free. Hardware makers that build it into devices, and model-as-a-service businesses with more than US$20 million of yearly revenue, need a separate license, and products with more than 100 million users or US$20 million of revenue in a month must show "Powered by StepFun AI".
Good for
- Self-hosting a large agent for coding, long documents and finance work with a 1M-token context
- Teams with an 8-GPU H200 or 4-GPU B300 node that want an open alternative to API-only agents
- Analyzing long reports, charts and screenshots, since it accepts images
- Research on a large open mixture of experts that ships FP8, BF16 and NVFP4 weights
Watch out for
- StepFun compared its High setting with its rivals' Max settings, and the scores are its own apart from GDPval-AA, so the margins are soft
- The custom license needs a separate agreement from hardware makers and from model-as-a-service businesses above US$20 million a year
- It needs a multi-GPU node even in NVFP4, at 394 GB
- The model card mentions video in one place and lists only text and image in its specification table
- It is a preview, so expect later versions to change
Performance claims come from the developer and have not been independently verified.
How much memory does it need?
Estimates for the Step-5-Preview (FP8) variant. A mixture-of-experts model must hold all of its experts in memory even though only some are used for each token, so total parameters decide the memory and active parameters decide the speed.
| Variants | Parameters | BF16 | FP8 / INT8 | 4-bit (Q4) | Fits on |
|---|---|---|---|---|---|
| Step-5-Preview (FP8) FP8 | 604B 27B active | 1.39 TB | 695 GB | 417 GB | Big-memory desktop |
| Step-5-Preview-BF16 BF16 | 604B 27B active | 1.39 TB | 695 GB | 417 GB | Big-memory desktop |
| Step-5-Preview-NVFP4 NVFP4 | 604B 27B active | 1.39 TB | 695 GB | 434 GB | Big-memory desktop |
Estimates: weights plus about 15% overhead. Real usage grows with context length and depends on the software.
The license in plain words
A custom license that lets you use, modify, distribute and sublicense Step 5 Preview free of charge, including commercially. Hardware makers that build the model into phones, vehicles, glasses, wearables, robots or smart-home devices need a separate commercial license, and so does any business that runs a model-as-a-service offering with more than US$20 million of yearly revenue. A product with over 100 million active users, or over US$20 million of revenue in a month, must show "Powered by StepFun AI". Models trained to a material extent on the model's outputs count as derivative works.
- Attribution or model name must be shown
- Free commercial use only below a revenue cap
- Separate deal needed to sell hosted access above a threshold
See also
- Best open-source LLMs in 2026
- Best open-source models for coding and AI agents
- Best open-source multimodal models in 2026 (images, video and audio)
Frequently asked questions
- What is Step 5 Preview?
- Step 5 Preview is a very large open model from the Chinese company StepFun, built for coding agents, long multi-step tasks and professional work such as finance. It has 600 billion parameters but uses 27 billion for each token, reads up to a million tokens and can look at images. Running it takes a multi-GPU server, and the license has special terms for hardware makers and hosting businesses.
- Can I use Step 5 Preview commercially?
- Yes, with conditions. Step 5 Preview is released under the StepFun 1.0, which allows commercial use but adds extra rules: Attribution or model name must be shown; Free commercial use only below a revenue cap; Separate deal needed to sell hosted access above a threshold. Read the license text before you ship a product.
- Is Step 5 Preview open source?
- Step 5 Preview is open-weight, not open source in the strict sense. You can download the model files, but the StepFun 1.0 adds conditions that the Open Source Initiative definition does not allow.
- Can I run Step 5 Preview on my own computer?
- Only on a machine with a lot of memory. The smallest variant, Step-5-Preview (FP8), needs about 417 GB at 4-bit precision. Desktops with 128 to 512 GB of unified memory can hold it, at modest speed.
- When was Step 5 Preview released?
- Step 5 Preview was released by StepFun on October 15, 2026.
- How long a context window does Step 5 Preview have?
- Step 5 Preview supports a context window of up to 1M tokens.
Official links
Sources
Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.
- Step 5 Preview (StepFun model page) stepfun.com
- Step-5-Preview model card (Hugging Face) huggingface.co
- StepFun Model License 1.0 (Step-5-Preview repository) huggingface.co
- Step-5-Preview-BF16 model card (Hugging Face) huggingface.co
- Step-5-Preview-NVFP4 model card (Hugging Face) huggingface.co
Release date and facts last verified November 11, 2026.
Related models
| Model | Released | Parameters | Context | License | Smallest variant needs |
|---|---|---|---|---|---|
| Step 3.7 Flash StepFun | 201B 11B active | 256K | Apache 2.0 (Permissive) | 120 GB Big-memory desktop | |
| Kimi K3.1 Pro Moonshot AI | 2.8T 104B active | 1M | Kimi K3 License (Conditional) | 1.72 TB Multi-node cluster | |
| GLM-5.3 Z.ai (Zhipu AI) | 753B 40B active | 1M | GLM-5.3 License (Conditional) | 520 GB GPU server | |
| DeepSeek V4 DeepSeek | 284B – 1.6T 49B active | 1M | MIT (Permissive) | 176 GB Big-memory desktop |