One million tokens and 600 billion parameters: StepFun opens the doors to Step 5

StepFun releases the weights of Step 5 Preview, a 600-billion-parameter AI model featuring a one-million-token context length and designed for agents.

Six hundred billion parameters, a context window of one million tokens, and the ability to run the model on one's own infrastructure. Step 5 Preview takes a new step forward with the release of its weights on Hugging Face and ModelScope. Already accessible via API since September, the flagship model from Chinese lab StepFun is now opening up to developers and enterprises looking to deploy their own AI agents, particularly in software development, finance, and consulting. One Million Tokens for Long-Running Tasks

Step 5 Preview relies on a Mixture-of-Experts (MoE) architecture totaling 600 billion parameters, with 27 billion activated for each processed token. This setup allows the model to mobilize only a portion of its parameters at each generation step, rather than the entire set.

Its context window can reach up to one million tokens—a volume designed for tasks requiring the examination of vast document sets, exploring code repositories, or maintaining history across numerous tool interactions.

The model accepts text and images as input, while its documentation also mentions video. However, the technical spec sheet details only text and image in its table of officially supported modalities. Outputs are generated in text format.

StepFun also announces an output capacity of up to one million tokens in the released version, though real-world usable limits depend on deployment configurations. From Software Development to Financial Analysis

The model specifically targets agents capable of working on long-duration tasks. In software engineering, it can navigate a project, modify multiple files, run tests, and resume its work after encountering an error.

StepFun also highlights the generation of structured reports from documents, spreadsheets, charts, and other business information.

Finance is among its specialized domains. On FrontierFinance, Step 5 Preview achieves 66.4 points, compared to 64.1 for GLM-5.3 and 62.6 for Kimi K3 in the comparison published by StepFun. However, Claude Opus 5 maintains a lead with 69.7 points.

In software engineering, Step 5 Preview scores 67.7 on DeepSWE v1.1, a result close to Kimi K3 and GLM-5.3, but below the 74.1 points awarded to GPT-6 Astra.

Performance varies depending on the task. On Terminal-Bench v4, for instance, Step 5 Preview scores 33.3 points, compared to 57.9 for GPT-6 Astra.

These figures come from the lab's technical presentation and results derived from various evaluation benchmarks. They help position the model without guaranteeing comparable performance across all production environments. Weights Released in Multiple Formats

The main news centers on the availability of the model files. StepFun notably offers downloadable FP8 and BF16 versions.

FP8 precision reduces memory requirements compared to BF16, at the cost of certain compute and compatibility constraints.

StepFun provides deployment configurations using vLLM, SGLang, and Transformers. In the documented examples, the FP8 version can run on four NVIDIA B300 GPUs or eight H200 GPUs. The BF16 version demands more resources—up to sixteen H200s in the presented configuration.

Therefore, this is not a model intended to run on a standard personal computer. Its release primarily enables organizations with significant GPU infrastructure to choose their hosting environment and inference tools. Open Weights, but a Specific License

The release is issued under the StepFun Model License 1.0, a license inspired by Apache 2.0 but subject to additional conditions.

Use, modification, and redistribution are permitted within the framework defined by this license. However, certain commercial activities require a separate agreement, particularly for hardware manufacturers integrating the model into devices or Model-as-a-Service providers exceeding specific revenue thresholds.

Attribution requirements may also apply to products reaching very high audiences or revenue figures.

Step 5 Preview is thus an open-weight model, but its license should not be equated with a standard Apache 2.0 license free of additional commercial restrictions. An Alternative to Proprietary APIs for Enterprise Agents

This release arrives amid a surge in high-capacity open models. Their appeal no longer rests solely on leaderboard performance, but also on the ability to customize their operation, control their hosting, and integrate them into internal systems.

For enterprises, however, this autonomy assumes having the necessary resources to run the model, as well as the expertise to manage and secure the infrastructure.

Step 5 Preview was already usable via API; opening its weights now alters deployment possibilities. Developers can inspect and adapt the model, while sufficiently equipped organizations can run it without relying exclusively on StepFun's hosted service—a technical freedom whose value will depend as much on licensing terms and operational costs as on the model's capabilities themselves.