As of October 15, 2026, businesses building document-analysis or coding assistants can use Step 5 Preview through StepFun’s paid API or run its downloadable model files on infrastructure they manage. The self-hosting option is a server deployment: published examples use multiple data-centre GPUs, not an office laptop. StepFun’s deployment guide.
For a team with uncertain demand and no model-serving staff, our analysis favours a limited API evaluation, provided account access and data terms fit. Self-hosting deserves a closer look when control over deployment or a fixed model version is a specific requirement. Downloadable weights alone do not establish a cost saving.
What a 100-job API bill looks like
StepFun lists three US-dollar rates per million tokens, the units of model input and output: $1 for uncached input, $0.10 for cached input and $2.70 for output. Output billing includes reasoning as well as the final answer. Current rate card.
Illustrative calculation: assume 100 jobs, one request each, with 100,000 input tokens and 10,000 total output tokens per request. These are chosen quantities, not a measured workload.
Scenario for those 100 jobs | Model-token charges |
|---|---|
All input uncached | $12.70 |
80% of input billed at the cache-hit rate; output unchanged | $5.50 |
All input uncached; output doubled to 20,000 tokens per job | $15.40 |
Baseline arithmetic: 10 million input tokens × $1 per million + 1 million output tokens × $2.70 per million = $12.70. These totals exclude extra calls, tools, storage, tax and human review. Pricing basis.
The cache scenario is hypothetical, not a promised discount. StepFun reuses matching prompt prefixes, but entries can expire and cache hits are not guaranteed. Check the reported cached-token count before budgeting around reuse. Prompt-cache documentation.
What running it yourself requires
FP8, BF16 and NVFP4 name different numerical formats for storing model weights. The published examples make the scale of the infrastructure choice clear:
FP8: four NVIDIA B300 GPUs or eight H200 GPUs in the vLLM serving recipe.
BF16: eight B300 GPUs, or sixteen H200 GPUs across two servers, in StepFun’s deployment examples.
NVFP4: four B300 GPUs in the vLLM recipe, with a configured context limit of 320,000 tokens. Lower precision does not automatically imply a smaller server package.
These are documented configurations, not proven minimums. The vLLM guide says available context and concurrency depend on remaining memory for request state. Its one-million-token context budget includes input and output together; operators may need to shorten requests or serve fewer at once. Serving constraints.
The hosted API separately lists a one-million-token context and a 64,000-token maximum output. Do not import the model card’s up-to-one-million-token output specification into an API purchasing decision. Hosted specifications; model-card specifications.
Software readiness matters too. SGLang says Step 5 support requires its preview container and has not yet shipped in a regular release. A deployment evaluation should pin the model revision and serving software, then check the application’s actual requests. SGLang’s current instructions.
Check which commercial-use rules apply
StepFun Model License 1.0 permits conditional no-charge use; it is not standard Apache 2.0. Its separate-agreement rules cover the model and derivatives:
Device manufacturers and brand owners: an agreement is required before commercializing capabilities in end-user devices or accompanying services, including remote delivery. This branch has no revenue threshold.
Model-service businesses: commercial model use requires an agreement if you or an affiliate operate a model-as-a-service business and combined all-source revenue exceeds $20 million over any consecutive twelve months.
Exceptions: section 2(c) excludes strictly internal use without third-party access to the model, its capabilities or outputs, and certified-channel use within the certification’s scope.
Model-as-a-service means third-party inference or fine-tuning access with meaningful control over inputs, parameters or training data. Specific end-user features without that control, and merely relaying another provider’s model, are excluded. Section 2(c).
Redistribution and large-scale attribution have additional conditions. Check the full licence against your use; do not assume a reseller is certified.
Compare the cost of a usable result
For self-hosting, request a quote that includes server rental or hardware depreciation, power where applicable, networking, storage, operations and recovery capacity. Match the quote to your chosen model format, request length, number of simultaneous users and required response time. Existing GPUs still have operating costs and could be needed for other work.
Our suggested comparison is cost per accepted job: total relevant costs over a fixed period divided by jobs that meet an agreed quality and turnaround standard. Include failed attempts and retries in costs, and treat shared application and review costs consistently on both sides. No infrastructure quote or measured throughput was established for this guide, so a self-hosting break-even point cannot be calculated.
Before committing, collect three pieces of evidence:
A defined task and acceptance standard. For document answers, check source references and critical figures; for code changes, define the tests and review needed. If the immediate problem is extracting text or layout from scans, the LightOnOCR-3 decision guide addresses that separate stage.
The full application workload. Count repeated calls and total generated tokens, not just the visible answer. StepFun says the integrating application supplies search, code execution and other tools; buying model access does not supply the whole assistant. Application responsibilities.
A documented data path. Confirm the provider’s retention, training-use and location terms before sending business material. Those protections were not established by the sources reviewed here. With self-hosting, include tools, logs and storage in the assessment; moving model inference alone does not make the entire workflow local.
Control and savings are different reasons to self-host. A team may accept higher operating costs to meet a deployment requirement. If the objective is lower cost, the deciding evidence is a comparable workload delivering acceptable results—not the fact that model files can be downloaded.
Methodology: This guide uses published StepFun documentation, the model licence and serving-framework instructions, checked October 15, 2026, with AI assistance. Calculations are illustrative. No model or API was run, and no performance, reliability or self-hosting savings were measured.