StepFun Releases Step 5 Preview Open Weights: 600B MoE, 27B Active, 1M Context
What changed
StepFun has released open weights for Step 5 Preview, its large-scale sparse Mixture-of-Experts (MoE) model. The full model includes 600 billion parameters, but only 27 billion are active per token during inference. It supports a huge context window of up to one million tokens. Both FP8 and BF16 precision checkpoints are available on Hugging Face and ModelScope, making it broadly accessible to developers.
Why builders should care
Step 5 Preview offers a rare combination of extreme scale, sparsity, and massive context length tailored for long-horizon tasks that require agentic reasoning. The sparsity design means less compute per token compared to a dense model with the same parameter count. Access to open weights in efficient FP8 and BF16 formats lowers barriers to experimentation for anyone aiming to build complex AI systems handling multi-step processes or long-term memory use cases.
The practical takeaway
Operators and builders can now trial a 27B-active parameter model that spans one million tokens context length at no cost. This enables work on advanced agent workflows, including multi-session dialogue, extended document understanding, and agentic control tasks that outstrip most existing open or commercial models. The open availability means startups and researchers do not have to depend on closed APIs or expensive proprietary engines to explore scaling benefits from sparse MoE architectures and ultra-long contexts.
What to watch next
The next steps will be in seeing how well Step 5 Preview performs on real tasks requiring planning and long-term dependencies and how effectively it operates in public infrastructure. Watch for updates on fine-tuning support, integration tools, and how the community adopts the ultra-long context window for practical applications. The sparse MoE design here could pressure rivals to expand their model scale or context windows while managing compute costs.
AI Quick Briefs Editorial Desk