StepFun, a Chinese AI lab, fulfilled the commitment announced in September and released today (October 15, 2026) the open weights of Step 5 Preview. The model, previously available only via API, can now be downloaded on Hugging Face and ModelScope in BF16, FP8 and NVFP4 versions.
What happened
In an official post on X, the company announced: "Step 5 Preview is now open weights." The checkpoints include a sparse Mixture-of-Experts model with 600 billion total parameters, of which about 27 billion are activated per token. It offers a context window of up to 1 million tokens, native text and vision input (with mentions of video), and is positioned as the flagship for long-horizon agentic work — especially software engineering, professional knowledge and finance.
The release includes deployment guides for vLLM, SGLang and Transformers, plus Day-0 support from platforms such as NVIDIA, Together AI and others. The model had already been available via StepFun’s API since September 20, with prices of about $1 per million input tokens and $2.70 output.
Why it matters
Large-scale open-weight models allow developers, researchers and companies to run, fine-tune and inspect frontier systems without relying exclusively on proprietary APIs. For the geek and development ecosystem, this expands options for local or self-hosted agents, especially useful in privacy, customization or controlled-cost scenarios — although the model size (about 1.2 TB in BF16) requires robust multi-GPU infrastructure.
StepFun joins a growing group of Chinese labs competing near the frontier with open or semi-open offerings, complementing the landscape dominated by OpenAI, Anthropic and Google.
Context
Step 5 Preview was introduced on September 20 as an advance on the "Pareto frontier" of intelligence versus cost. The promise of open weights on October 15 was kept, according to the official post and public repositories. The license is the StepFun Model License 1.0, which allows use, modification and redistribution with some commercial conditions for large-scale uses.
Sources
Transparency: This content was created, edited or reviewed with the assistance of artificial intelligence. The information was cross-checked with public posts on X and sources available on the internet. Consult the original sources to verify the full context.
By GeekikiBot
