News · SI
StepFun has published the weights for Step 5 Preview, a 600B-parameter mixture-of-experts model with 27B active and a 1M-token context, built for coding and long-running agent work. Here is what it takes to run it and what the licence lets you do.
StepFun has published the weights for Step 5 Preview, its flagship model for agent work, on Hugging Face and ModelScope[^1]. It is a sparse mixture-of-experts model with 600 billion total parameters, 27 billion active per token and up to 1 million tokens of context, and it takes text, images and video as input[^1][^2].
01 / 05
StepFun ships FP8, BF16 and NVFP4 builds, with ready-made recipes for vLLM, SGLang and a standalone Transformers implementation[^1]. The smallest setup on the model card is the FP8 checkpoint on one node of four B300 GPUs; on Hopper it needs eight H200s, and BF16 doubles that to two H200 nodes[^1]. StepFun warns that even these configurations do not guarantee memory for a full 1M-token request once the KV cache is counted[^1]. Speculative decoding through multi-token prediction and dedicated reasoning and tool-call parsers are built into the serving commands[^1].
02 / 05
On StepFun's own chart, Step 5 Preview on its High reasoning setting leads Kimi K3 and GLM-5.3 on DeepSWE v1.1 (67.7), ProgramBench (80.5) and FrontierFinance (66.4), but trails GPT-6 Astra and Claude Opus 5 on most coding tests, and GLM-5.3 on Terminal-Bench v4 (33.3 against 41.9)[^1]. Those are StepFun's numbers, except its GDPval-AA score, which comes from Artificial Analysis[^1]. Treat it as a strong open model rather than a frontier replacement, and test it on your own repositories.
03 / 05
The same model runs on StepFun's API as step-5-preview at US$1.00 per million input tokens, US$0.10 on a cache hit and US$2.70 per million output tokens[^3]. The hosted version accepts up to 1M tokens in and 64K out, and supports tool calling, JSON Schema output, prompt caching and three reasoning-effort levels[^2].
04 / 05
The StepFun Model License 1.0 is Apache-2.0 style with extra conditions[^4]. Very large products must credit "Powered by StepFun AI", and device makers or model-hosting businesses above US$20 million in yearly revenue need a separate commercial licence[^4].
05 / 05
For most small teams the weights matter less than the option they create. You can build an agent against the cheap hosted API today and know that, if data residency or cost later forces you to self-host, the same model is downloadable and runs on standard serving stacks. If you sell hardware, especially anything in a car, a phone or a robot, talk to StepFun before you build on it.
FAQ
Step 5 Preview is StepFun's flagship model for agentic work. It is a sparse mixture-of-experts model with 600 billion total parameters and 27 billion active per token, a context window of up to 1 million tokens, and text, image and video input. StepFun says it is aimed at software engineering, long-running agent tasks and professional knowledge work, with particular strength in finance.
StepFun published the weights on Hugging Face and ModelScope on 15 October 2026. There are three checkpoints on Hugging Face: FP8 (stepfun-ai/Step-5-Preview), BF16 (stepfun-ai/Step-5-Preview-BF16) and NVFP4 for Blackwell GPUs (stepfun-ai/Step-5-Preview-NVFP4).
This is data centre hardware, not a desktop model. StepFun's model card lists the FP8 checkpoint on one node of 4 B300 GPUs or 8 H200 GPUs with vLLM or SGLang, and BF16 on 8 B300 GPUs or two nodes of 8 H200 GPUs. It also warns that these setups do not guarantee enough memory for a full 1M-token request, because the KV cache needs room too.
On the StepFun Open Platform, step-5-preview costs US$1.00 per million input tokens on a cache miss, US$0.10 on a cache hit, and US$2.70 per million output tokens, with reasoning tokens counted as output. The hosted API caps output at 64K tokens.
Mostly yes. The StepFun Model License 1.0 is Apache-2.0 style, but it adds conditions. Products with more than 100 million monthly active users or more than US$20 million in monthly revenue must display 'Powered by StepFun AI'. Hardware makers shipping it in devices such as phones, cars or robots, and model-as-a-service businesses with more than US$20 million in yearly revenue, need a separate commercial licence.
That’s the story
Spectrum Web Co ships SI and Web3 products, places talent and builds social narratives as monthly services.
