StepFun’s Step 5 Preview is now available as open weights, giving developers a model aimed at coding agents and other long-running tasks that they can download and self-host. The headline trade-off is substantial capability on paper, but a data-centre-sized appetite for hardware.
StepFun Watch analysis
What happened
A MarkTechPost report published on 17 October says FP8 and BF16 checkpoints are available through Hugging Face and ModelScope. It describes Step 5 Preview as a sparse mixture-of-experts model with 600 billion parameters in total, 27 billion active per token and a context window of up to one million tokens. The report says the model is intended for coding, document analysis and finance workflows.
Self-hosting is not a casual weekend download. The report lists an FP8 setup using four B300 GPUs or eight H200s, and says the FP8 checkpoint is about 650GB. It also reports an API price of $1 per million input tokens and $2.70 per million output tokens.
The reported results are mixed rather than a clean victory lap: StepFun’s own evaluations include 33.3% on Terminal-Bench 4.0, while Artificial Analysis reports 88.3% on its long-context AA-LCR test. The article notes that StepFun and rival models were evaluated at different reasoning settings, so those comparisons are not like for like. Read MarkTechPost’s report.
Why it matters
Open weights give developers more control than an API alone: they can run the model on their own infrastructure and investigate how it performs for their work. But “open” does not mean “easy to run”. The listed GPU requirements put self-hosting firmly in the data-centre bracket, while the API offers a lower-friction route with a bill attached.
The million-token context window and agentic-work focus make this a notable new option for teams handling large codebases or documents. The benchmark caveat matters too: a long context window is useful, but it does not settle how well the model performs across everyday tasks.
Our read
Step 5 Preview is worth watching for teams with serious compute and a reason to keep model weights in-house. For everyone else, the API is the practical doorway, and independent, matched evaluations will tell us more than a tall parameter count ever could. That is a rather expensive way to learn whether a model suits your workflow.
What to watch
- Whether independent evaluators publish matched comparisons with other open-weight models.
- How well Step 5 performs on real coding and document-analysis workloads.
- Whether serving tools and hardware options make self-hosting more accessible.
Discussion spark: Would you choose a model with downloadable weights and a million-token context window if self-hosting required data-centre hardware, or is API access enough?
Sources and evidence
- StepFun Releases Step 5 Preview Open Weights: 600B MoE, 27B Active, 1M Context – MarkTechPost (17 October 2026, 06:29 UTC)
not affiliated with or endorsed by StepFun