StepFun has released Step 5 Preview as open weights, providing developers with a powerful new self-hostable option for long-horizon agentic work, coding agents, document analysis, and finance workflows. The sparse Mixture-of-Experts model features 600 billion total parameters, 27 billion active parameters per token, and an expansive context window of up to 1 million tokens. FP8 and BF16 checkpoints are now live on Hugging Face and ModelScope, alongside a managed API.
Designed for professional knowledge work, the model navigates large codebases, runs code, iterates on results, and analyzes long documents, charts, screenshots, and video. Its architecture includes 92 transformer blocks in a narrow, deep design, utilizing gated sliding-window attention, sparse grouped-query attention, speculative decoding, and FP8 mixed precision training. Independent evaluations from Artificial Analysis highlight strong reasoning capabilities, including an 88.3% score on AA-LCR long-context reasoning and 46.5% on Humanity's Last Exam.
Running the model requires substantial data-center hardware due to its massive scale. The FP8 checkpoint demands at least four B300 GPUs or eight H200 GPUs, while the BF16 version requires eight B300 GPUs or 16 H200 GPUs. Why it matters: This release gives enterprises and researchers a top-tier open-weight model capable of advanced reasoning and long-context agentic tasks.