What was released
StepFun shared open weights for Step 5 Preview, its flagship model built for agentic work, on Hugging Face and ModelScope. FP8 and BF16 checkpoints are available. The API launched on September 20, with weights following on October 14.
Architecture
Step 5 Preview is a sparse Mixture-of-Experts model: only 27B of its 600B parameters activate per token. It uses 92 transformer blocks in a narrow, deep design. Attention layers interleave at a 3:1 ratio, with three gated sliding-window blocks for every full-attention block. Full-attention layers use sparse GQA that picks 4,096 history tokens per query, reaching up to 1M tokens of context. An MTP-3 head adds roughly 30% end-to-end speedup.
Hardware needs
The model runs only on data-center GPUs. The FP8 version needs 4x B300 or 8x H200; BF16 needs 8x B300 or 16x H200. StepFun reports the FP8 checkpoint at 650.32 GB.
Performance
In StepFun's own evaluations, the model scores 67.7 on DeepSWE v1.1, 33.3 on Terminal-Bench 4.0 and 66.4 on FrontierFinance. Artificial Analysis data shows 88.3% on AA-LCR long-context reasoning, 46.5% on Humanity's Last Exam and 43.7 on the Intelligence Index. StepFun's chart ran rivals at different effort settings, so the comparison is not like-for-like. Vals AI results reveal weak spots: 0.4% on SRE Bench and 9.2% on Harvey's Legal Agent Benchmark.
