Step 5 Preview Weights Arrive With Commercial Limits
The 600B model now has documented self-hosting paths, but million-token requests and a custom license complicate deployment and commercial use.
Listen
AI narration
11:08
0:00 / 11:08
AI SummaryGenerated from this article
StepFun released downloadable weights for its 600-billion-parameter Step 5 Preview model with documented self-hosting paths for vLLM, SGLang and Transformers, supporting up to one million tokens of combined input and output. The sparse mixture-of-experts model requires substantial GPU infrastructure, starting at four B300 GPUs for FP8 quantization, but operates under a custom license with commercial restrictions that differ from standard open-source terms, including attribution requirements for products exceeding 100 million users or $20 million in revenue and separate agreements for hardware makers and model-as-a-service operators earning over $20 million annually.
StepFun has made the Step 5 Preview weights downloadable, giving teams a self-hosting option for its 600-billion-parameter model. The release includes documented deployment paths for vLLM, SGLang and Transformers, alongside support for a context window of up to one million tokens.
The October 15 weights release delivers the expansion in access promised after September’s API preview. It does not introduce a new model architecture.
Builders have two questions to work through: what infrastructure the model needs, and what its license permits. StepFun grants broad no-charge usage rights, but its custom terms include attribution requirements and separate-commercial-license triggers that differ from ordinary Apache 2.0 licensing.
One Million Tokens Is a Shared Input-and-Output Budget
According to the Step 5 Preview repository, the model uses a sparse mixture-of-experts architecture with 600 billion total parameters and 27 billion active parameters. Activating a subset of parameters during inference does not make this equivalent to hosting a dense 27B model. The deployment still has to accommodate the much larger collection of weights.
StepFun positions the model for repository-level software engineering, repeated tool use and professional knowledge work, including financial analysis. These are vendor descriptions of its intended capabilities, not guarantees about performance in a particular deployment.
Although the specifications list both context and maximum output as “up to 1M tokens,” the serving instructions explicitly state that the limit covers input and output combined. A request that fills most of the window with documents, code and accumulated tool responses leaves less room for generation. It does not receive a million-token input allowance plus another million tokens of output.
That shared budget could be useful for agents that need to retain information across successive tool calls. A large configured window alone, however, does not establish reliable retrieval across that window, successful task completion or affordable latency. Those remain workload-specific evaluation questions.
Work with Zeniteq
Let’s work together
We’re open to thoughtful collaborations with teams building in AI. Explore the ways we can work together.
The release includes concrete infrastructure configurations. StepFun’s deployment matrix lists these vLLM examples:
Checkpoint precision
Documented hardware configuration
FP8
One node with four B300 GPUs
FP8
One node with eight H200 GPUs
BF16
One node with eight B300 GPUs
BF16
Two nodes with eight H200 GPUs each
These are documented configurations, not a universal minimum-hardware specification or a promise that every listed setup can sustain million-token requests at useful concurrency.
The repository explicitly warns that actual memory requirements, usable context length and concurrency depend on the checkpoint, framework and memory allocated to the key-value cache. That cache holds attention-related state during inference and needs memory alongside the weights and runtime buffers.
Loading the model successfully is only the first deployment test. A team may still need to reduce context length or concurrent requests to stay within available memory.
The framework paths serve different needs:
vLLM provides OpenAI-compatible API serving, with FP8 and BF16 examples.
SGLang also provides OpenAI-compatible serving, with documented FP8 configurations.
Transformers provides a standalone offline-generation path, including FP8 and BF16 instructions.
StepFun tells users to choose framework builds that include Step 5 Preview support, including the reasoning, tool-call and speculative-decoding features used in its commands. The vLLM examples enable automatic tool choice and Step 5-specific reasoning and tool-call parsers.
This provides meaningful integration support. An application still needs a complete agent runtime to execute tools, apply permissions and handle failures; tool-call parsing does not supply those functions.
There are two additional BF16 caveats. Setting --dtype bfloat16 on an FP8 checkpoint does not convert its quantized weights into a BF16 checkpoint. The repository also warns that the full BF16 conversion-to-generation workflow has not yet been validated end to end. Teams choosing that route should plan to validate it themselves, with no finished production assurance.
Attribution and Commercial Licensing Have Different Triggers
The governing document is the StepFun Model License 1.0. It grants no-charge, royalty-free rights to use, copy, modify and redistribute the model, subject to its conditions.
Commercial use is possible under the public license, with three provisions worth examining separately.
Large Products Must Display Attribution
If a commercial product or service using the model exceeds either of these thresholds in any calendar month, the license requires attribution thereafter:
More than 100 million active users.
More than US$20 million in revenue.
The required wording is “Powered by StepFun AI,” unless StepFun designates another model name in writing. It must appear prominently in the product interface, including the About page, and in its user agreement or product documentation, with the specified prominence relative to other third-party component notices.
Crossing this attribution threshold does not, by itself, mean the product must negotiate a separate commercial agreement. The license sets out different conditions for that.
Hardware Makers Need an Agreement Before Commercial Availability
Manufacturers or brand owners making the model’s capabilities commercially available as a feature of end-user hardware must first enter a separate commercial agreement.
The clause covers devices such as phones, vehicles, glasses, wearables, robots and smart-home hardware. It is not limited to embedding weights on the device: it also reaches capabilities supplied remotely or through accompanying services, whether available at shipment or introduced later through activation, downloads or updates. Routing implementation through a third party does not automatically avoid the provision.
There is no revenue threshold attached to this hardware condition. A small hardware company should therefore not assume it qualifies for the same treatment as a small standalone software application.
MaaS Operators Face an Annual Revenue Test
A separate agreement is also required for commercial use when the user or an affiliate operates a qualifying model-as-a-service business and their combined revenue from all sources exceeds US$20 million over any consecutive twelve-month period.
Unlike the attribution rule, this test uses annual revenue, includes affiliates and counts revenue from all sources, including revenue unrelated to Step 5 Preview.
The license defines MaaS as giving third parties inference or fine-tuning access to the model or a derivative, including through an API, where those third parties have meaningful control over inputs, parameters or training data.
Merely relaying requests to another provider’s model is excluded. The license also distinguishes a specific end-user feature that does not give customers that control, though such a feature could still fall under the separate hardware provision.
The commercial-agreement requirements include exceptions for strictly internal use that does not expose the model, outputs or underlying capabilities to third parties, and for use through StepFun-certified distribution channels within their certification scope. These exceptions apply to that section, not to every license obligation.
Fine-Tuning Does Not Remove the Conditions
The license’s derivative-work definition covers fine-tuned, distilled, merged, quantized, pruned and otherwise modified models containing or reproducing a substantial part of the original.
It also treats a model trained to a material extent on Step 5 Preview outputs as reproducing or containing a substantial part of it. Teams considering synthetic-data training should examine that language; coverage is not limited to redistributed original weights.
An application does not become a derivative work solely because it calls the model through an API. Conditions that separately apply to products using the model can still apply.
Redistribution brings further obligations: include the license, preserve required notices, identify modifications prominently, include any supplied NOTICE file, and provide the model card or a link to it.
The accurate description is open weights under a custom license. Treating this release as ordinary Apache 2.0 weights would hide commercially consequential differences.
AI Pricing Now Includes Infrastructure and Contract Questions
For teams comparing AI pricing, no-charge weight access answers only one part of the cost question. Self-hosting replaces dependence on a hosted endpoint with responsibility for GPU capacity, serving operations and workload validation.
The public license also specifies when a separate agreement is necessary without publishing a price for that agreement. An affected hardware vendor or MaaS operator cannot calculate its licensing cost from the repository alone.
Teams should assess technical fit and license eligibility in parallel: test the required context length, concurrency and tool-use workflow on the intended hardware while checking the actual business model against the license. Neither a successful model load nor a small initial deployment resolves both questions.
StepFun’s release creates a documented path to operating Step 5 Preview outside its hosted API. Its value to a particular team depends on whether that team can run it economically and use it under acceptable terms.
