Evidence at a glance
The Weights Are Open; the Hardware Bar Remains
StepFun has released open weights for Step 5 Preview, a model aimed at agentic work in software engineering, professional knowledge work, and finance. FP8 and BF16 checkpoints are available on Hugging Face and ModelScope. The company launched an API on September 20 and followed with weights on October 14. The model has 600 billion total parameters, activates about 27 billion per token, and supports a context window of up to one million tokens.
Those figures can invite a misleading inference: if only about 4.5% of the parameters are active at a time, deployment must be relatively light. But active parameters describe which weights participate in processing a token, not the storage required for the full model, and certainly not whether an ordinary server can host it. The stated minimum is four B300 or eight H200 GPUs for FP8, and eight B300 or sixteen H200 GPUs for BF16. The FP8 checkpoint is about 650.32 GB. Open weights give teams a self-hosting option, but they do not remove the data-center hardware threshold.
How Sparse Routing Meets Long Context
Step 5 Preview does not apply the same attention operation to the entire history at every layer. Its 92-layer, deep-and-narrow design interleaves three gated sliding-window attention layers with one full sparse-attention layer in a 3:1 pattern. Local layers focus on nearby content. Global layers use Sparse GQA to select information from the longer history instead of calculating every token-to-token relationship. StepFun says each global attention layer selects 4,096 history tokens per query and scores them in groups of eight.
The design targets two competing costs: long documents require access to distant information, while full attention can be expensive. The model also includes an MTP-3 head, which StepFun reports delivers roughly a 30% end-to-end speedup. This is not a promise that million-token context comes at no cost. The practical benefit depends on whether a task needs long-range retrieval, how inference hardware and serving software are configured, and whether the reported speedup holds for the workload in question. The architecture offers a way to reduce some costs; it does not replace measuring latency and throughput in the intended environment.
What the Scores Show—and What They Do Not
The reported results do not form a uniformly strong capability profile. Artificial Analysis gives Step 5 Preview 88.3% on its AA-LCR long-context reasoning evaluation, while the model scores 33.3% on Terminal-Bench 4.0, below GLM-5.3 at 41.9%. StepFun reports 67.7 on DeepSWE v1.1, close to Kimi K3 at 67.5 and GLM-5.3 at 66.9 in the same comparison table. On FrontierFinance, StepFun reports 66.4%, ahead of GLM-5.3 at 64.1 and Kimi K3 at 62.6, but behind Claude Opus 5 at 69.7%. These comparisons are better read as a task profile than as a single overall ranking.
Evaluation conditions also limit what can be concluded. The DeepSWE, Terminal-Bench, and FrontierFinance comparison figures are reported by StepFun. The material says Step 5 Preview was run at High reasoning, while the competitors in the chart used Max effort, so the comparison is not strictly like-for-like. Independent evaluations report 88.3% on AA-LCR and 46.5% on Humanity’s Last Exam, but only 0.4% on SRE Bench and 9.2% on Harvey’s Legal Agent Benchmark. The spread is a reminder that long-context performance does not automatically translate into command-line, operations, or legal capability. Benchmark scores also cannot replace checking tool-use failures, result verification, and end-to-end task completion.
Start a Self-Hosting Decision with the Workload
For technical leaders, the practical value of Step 5 Preview depends on whether the workload and deployment conditions match. If a task must integrate information across files in a large codebase or across sections of long documents, its million-token context and sparse attention make it a candidate for evaluation. If reliable terminal execution is the priority, the Terminal-Bench result is a concrete reason to test the model rather than a peripheral score to set aside. The finance result describes performance on a particular benchmark; it does not substitute for testing against a team’s own data and processes.
A deployment assessment should consider weight format, GPU capacity, inference latency, concurrency, and licensing together. The material identifies the license as StepFun Model License 1.0, but does not interpret its terms in enough detail to establish specific commercial restrictions. “Open weights” alone is not evidence that use is unrestricted. A prudent approach is to test whether long context produces measurable quality gains on target tasks, estimate cost and throughput with the intended serving configuration, and separately review the license and failure-handling requirements. If the gains do not justify the hardware and operational costs, open weights by themselves are not a reason to migrate.