Step 5 Preview open weights explained for deployment teams
StepFun’s Step 5 Preview offers downloadable weights for teams that want to operate its agent model themselves. The practical question is not just whether the files are available. It is which checkpoint fits the intended hardware, which serving software supports it and whether the planned business falls within the license.
The official repositories currently list FP8 files totaling 650 GB, BF16 files totaling 1.21 TB and a 394 GB NVFP4 repository. These are repository sizes, not promises about the GPU memory needed to run an application. The file listings contain actual weight shards rather than empty announcement pages.
This is a deployment guide to an existing release. StepFun’s technical account is dated October 15, 2026. The NVFP4 history records its weight upload on October 14. The repositories and serving instructions were checked on October 23, not treated as a new launch that day.
What hardware does Step 5 Preview need
The official deployment table gives FP8 examples using four NVIDIA B300 GPUs or eight H200 GPUs. BF16 examples use eight B300 GPUs or two H200 nodes with eight GPUs each. Those are starting configurations, not guaranteed capacity for every request.
StepFun lists 600 billion total parameters and 27 billion active parameters. The smaller active count does not mean only that portion needs storage. Its context allowance covers input and output together. Cache memory, runtime buffers and concurrent requests can reduce usable capacity below the advertised million tokens.
The documentation also distinguishes a genuine BF16 checkpoint from a runtime dtype setting. Changing that setting on FP8 files does not undo their quantization. Its standalone Transformers workflow for full BF16 shard conversion followed by generation is not yet validated from start to finish. That caveat does not say every BF16 execution path is untested.
FP8 and NVFP4 have different serving constraints
The SGLang cookbook still directs users to a dedicated preview Docker image. A standard package installation does not yet include the required support. Its FP8 recipes distinguish low latency for one user from throughput across multiple users, and explicitly say four H200 GPUs cannot hold the FP8 weights.
The same guide now includes NVFP4 for Blackwell hardware. It is not available on H200 in that guide and does not support the MegaMoE option. Smaller checkpoint files therefore do not make every existing GPU configuration compatible.
Cache precision is another independent choice. SGLang calls its optional FP8 cache lossy and recommends checking accuracy again. Teams should test representative documents, tool responses and concurrency rather than assuming a smaller cache preserves the same answers.
Why StepFun chose a deep architecture
StepFun’s October 15 architecture report describes 92 transformer blocks and a mixture of sliding window and sparse attention. The intended workload repeatedly reads long histories and invokes tools. This explains the design emphasis, but is not evidence that every agent application benefits equally.
The company says its Sparse GQA design groups historical tokens into blocks of eight, reducing indexing and selection cost by a factor of eight. That is a claim about one component, not an eightfold application speedup. The report also discusses instability during training and changes to precision and gradient handling. Those disclosures matter more to a deployment assessment than an unqualified claim of frontier performance.
What the commercial license allows and restricts
The StepFun Model License 1.0 permits use and modification under conditions. Redistribution requires the license and notices, identification of modifications and the model card or a link to it. Downloadable weights do not mean unrestricted commercial rights.
A separate agreement is required when a hardware manufacturer or brand owner commercially supplies the model’s capabilities through specified end user devices or accompanying services. Another provision applies to commercial use by a business or affiliate providing model inference or fine tuning access with meaningful customer control when combined revenue from all sources exceeds $20 million over twelve consecutive months. Exceptions cover defined internal use and certified distribution channels within their certification scope.
A separate attribution provision applies when a commercial product or service exceeds 100 million active users or $20 million revenue in a calendar month. Those attribution thresholds should not be confused with the commercial agreement test. Product teams should map their actual delivery arrangement to the full terms before committing to distribution.
A useful evaluation starts with the workload
The first decision is whether the project needs direct operation of the model enough to justify its infrastructure and maintenance demands. Record the selected checkpoint, serving image and hardware before comparing results. Then measure completion quality on real tasks alongside response latency and the longest requests the service must accept.
ByteForward’s vLLM serving report provides related context on separating memory behavior, request handling and application compatibility. It should not be read as a claim that a general vLLM release replaces StepFun’s preview instructions.
For this model, a sensible go or no go decision combines three checks. The selected configuration must run the intended workload, its output must meet the application’s accuracy requirements and its license must fit the distribution plan. This report reviews published documentation and files. ByteForward has not deployed Step 5 Preview or reproduced StepFun’s performance claims.
Custom editorial illustration created for ByteForward using StepFun branding.
