Model release

StepFun opens Step 5 Preview's 600B MoE weights with a custom commercial-use license

StepFun's downloadable multimodal MoE targets long-horizon agents, but its million-token context claim and Apache-style license language require precise qualification.

Open Model Weights published19 Oct 2026
Primary sourceStepFun official Step 5 Preview BF16 model card and license
Source published2026-10-15
Original technical illustration showing a Step 5 Preview 600B mixture-of-experts router, 27B active path and separate licensed checkpoint packages
Conceptual illustration of Step 5 Preview's published MoE scale and precision variants; not the actual network topology, a legal opinion or a benchmark measurement. AI-generated editorial illustration; not a documentary photograph.

StepFun has published official downloadable Step 5 Preview checkpoints after describing the model as a flagship for coding, agents and professional knowledge work. Its primary Hugging Face cards identify a sparse mixture-of-experts architecture with 600 billion total parameters, 27 billion activated per token and a context window of up to one million tokens. The published repositories include distinct precision variants rather than a single universally deployable image, and StepFun documents vLLM, SGLang and Transformers-based paths for inference.

The headline numbers need careful interpretation. Active parameters describe the selected computation path, not the amount of model data that must be stored or accessed. FP8, BF16 and NVFP4 packages have different serving characteristics and hardware requirements. A one-million-token supported window is a stated model capability, not a claim that every deployment can afford maximum-context attention, KV cache, multimodal preprocessing and useful throughput. Operators should calculate their own total memory and batching envelope before selecting a variant.

StepFun describes long-horizon tool workflows, software engineering, document analysis and finance as target use cases. The card attributes results to specified evaluation configurations, including a High reasoning setting and named benchmark harnesses. No comparative model test was run for this article. Benchmarks created or summarized by the publisher are helpful leads for evaluation, not independent proof of superior production agents.

The license is especially consequential. The actual weights are governed by StepFun Model License 1.0, not standard Apache 2.0. The text requires additional prominence and attribution after certain very large usage or revenue thresholds and a separate commercial agreement for specified hardware-device products and sufficiently large model-as-a-service businesses. Exceptions and definitions matter, so organizations should read the original grant and conditions before deploying, redistributing or building a commercial product. Step 5 Preview is a meaningful open-weight release, but open weights do not mean unlimited rights or inexpensive local operation.

OMW REGISTRY WATCH

What this changes for the evidence layer

Treat FP8, BF16 and NVFP4 official Step 5 Preview artifacts as separate source checkpoints if later considered for model intake. Inspect repository manifests, SHA-pinned weights, StepFun Model License 1.0 and variant-specific inference settings before any registry verification. Claimed 1M context and coding/finance benchmark results are publisher statements, not Open Model Weights load tests.

PRIMARY SOURCE

StepFun official Step 5 Preview BF16 model card and license

This brief is based on the cited primary source. Performance, benchmark and comparative claims remain attributed unless Open Model Weights publishes an independent measurement.

Open source article ↗