Step 5 Preview Open Weights: Hardware, Context and License Explained

StepFun's October 15 release of Step 5 Preview's weights gives developers a different decision to make. The model was already available through StepFun's products and API at its September 20 public announcement. The new milestone brings official BF16 and FP8 checkpoints to Hugging Face and ModelScope. Teams can now evaluate operating the model themselves, alongside the hosted option.

That distinction matters more than another round of launch benchmarks. Downloadable weights let an infrastructure team control the serving environment and investigate how the model behaves under its own workload. They also move responsibility for capacity, reliability, security, and operating cost onto that team. Step 5 makes this trade particularly clear because its relatively small active parameter count sits inside a much larger model.

A small active slice still needs a large server

StepFun describes Step 5 Preview as a sparse mixture-of-experts model with 600 billion total parameters and 27 billion active per token. Its context window reaches one million tokens. Sparse routing reduces the portion of the network used for each token, but the inactive experts still belong to the deployed model. Treating it like an ordinary 27-billion-parameter download would lead to a very different hardware estimate.

The official deployment guide gives FP8 examples using four B300 GPUs or eight H200 GPUs. BF16 examples use eight B300s or sixteen H200s across two nodes. It includes vLLM, SGLang, and Transformers deployment paths, but says the full BF16 Transformers shard-conversion and generation workflow is not yet validated end to end. The guide also warns that loading the weights does not guarantee sufficient memory for a million-token request: the KV cache and runtime buffers need room too. Input and output share the context budget.

For a team considering this release, the practical starting point is its expected workload: typical prompt length, simultaneous sessions, and acceptable response time. A configuration that handles one long request may have very different economics when several agents are running together. Those are measurements to make before turning the maximum context length into a product promise.

The architecture explains the ambition

StepFun's October 15 technical article describes a 92-block Transformer that alternates three sliding-window attention layers with one sparse full-attention layer. Its Sparse GQA mechanism uses block-wise token merging to reduce the work involved in selecting relevant history. The company frames these choices around long, tool-heavy tasks, where agents repeatedly ingest execution results and must reason over an expanding record of work.

The training discussion is equally useful. StepFun says it trains context compaction as part of the task, allowing an agent to reorganize its state and continue when the working context fills up. That is a more informative design choice than the context-window number alone. A long-running agent needs a way to preserve the right information after it runs out of room. Whether this particular approach preserves what your application needs still requires testing.

Read the license before planning a service

The release uses the StepFun Model License 1.0, which permits use, modification, and redistribution subject to its conditions. Hardware manufacturers or brand owners shipping the model's capabilities in devices or accompanying services need a separate commercial license. So do qualifying model-as-a-service businesses when combined revenue from all sources across the business and its affiliates exceeds $20 million over a consecutive twelve-month period. The commercial-license section includes an internal-use exception.

The license also treats models trained materially on Step 5 outputs as derivative works. That deserves attention from teams considering distillation or synthetic training data. Calling the release simply “open source” would leave out conditions that can affect how a company builds and distributes a product.

There is independent performance context, although it does not settle the self-hosting question. Artificial Analysis currently reports an Intelligence Index score of 44 and describes the model as very verbose. Its hosted-model measurements are useful for shortlisting, but they cannot establish throughput or cost for your own hardware and serving configuration. Track successful tasks, retries, and total generated tokens when comparing options.

The useful outcome of this release is a wider set of deployment choices backed by inspectable artifacts. Teams with substantial inference infrastructure now have something concrete to evaluate. Everyone else can still assess the model through hosted access before deciding whether operating it themselves would solve a real problem.

I’m Paras. I work at the intersection of AI, engineering and distribution.

I help technical founders turn product launches and complex engineering updates into clear stories that developers and founders want to read and share.

Book a 30-minute call.