StepFun Releases Step 5 Preview Open Weights for Agentic Work

StepFun describes Step 5 Preview as its flagship model for agentic work. It is intended to navigate large codebases, coordinate multi-file changes, run code and iterate on results. It can also analyze long documents, charts, screenshots and video to produce structured reports. The model is a data-center deployment only: the FP8 checkpoint alone is 650.32 GB. Listed configurations put FP8 on vLLM or SGLang on four B300 GPUs or eight H200 GPUs on one node; BF16 on vLLM requires one node with eight B300 GPUs or two nodes with 16 H200 GPUs. Transformers support is for offline generation only, and the full BF16 shard workflow has not been validated end to end, the report said. vLLM and SGLang provide dedicated Docker images with OpenAI-compatible APIs, and on Blackwell expert-parallel modes use the MegaMoE backend. StepFun recommends temperature 1.0, top_p 0.95 and top_k 0.

Architecturally, only 27 billion of the 600 billion parameters activate per token. StepFun’s technical team describes 92 transformer blocks in a narrow, deep design, with gated sliding-window attention and full attention layers interleaved at a 3:1 ratio. Full-attention layers use sparse grouped-query attention, selecting 4,096 history tokens per query and scoring them in blocks of 8. An MTP-3 speculative decoding head provides roughly 30% end-to-end speedup. The model was trained with FP8 mixed precision and reinforcement learning on million-level verifiable agentic environments.

MarkTechPost reported that StepFun ran its own evaluations at the High reasoning setting, while rivals in its comparison chart ran at Max effort, so the comparison is not like-for-like. On the Hugging Face card, Step 5 Preview scores 67.7 on DeepSWE v1.1, 33.3 on Terminal-Bench 4.0 and 29.5 on Agents’ Last Exam (ALE-CLI). Artificial Analysis numbers via OpenRouter put the model at 43.7 on the Intelligence Index, 88.3% on AA-LCR long-context reasoning, 46.5% on Humanity’s Last Exam and 54.3% on GDPval-AA.

The report also cited Vals AI results showing weaker spots: 0.4% on SRE Bench and 9.2% on Harvey’s Legal Agent Benchmark. On finance, StepFun’s chart shows 66.4% on FrontierFinance, above GLM-5.3 at 64.1% and Kimi K3 at 62.6%, but below Claude Opus 5 at 69.7%. In StepFun-reported comparisons, DeepSWE v1.1 is 67.7 for Step 5 Preview, 66.9 for GLM-5.3 and 67.5 for Kimi K3; Terminal-Bench 4.0 is 33.3 for Step 5 Preview, 41.9 for GLM-5.3 and 12.6 for Kimi K3.

The comparison with other open-weight models lists 600 billion total parameters for Step 5 Preview, 753 billion as GLM-5.3’s Hugging Face file size and 2.8 trillion for Kimi K3. Step 5 Preview and Kimi K3 list 27 billion and 104 billion active parameters respectively; GLM-5.3 does not disclose its active count. All three list 1 million-token context. Step 5 Preview accepts text and image input, with video added through the API; GLM-5.3 is text only; Kimi K3 accepts text and image. Weight formats are FP8 and BF16 for Step 5 Preview, BF16 and FP8 for GLM-5.3 and MXFP4 for Kimi K3. Step 5 Preview is released under the StepFun Model License 1.0. The minimum listed self-host setup for Step 5 Preview is four B300 GPUs with FP8; the other two models do not disclose a minimum setup in the comparison.

The StepFun API charges $1.00 per 1 million input tokens and $2.70 per 1 million output tokens, with cache hits at $0.10 per 1 million tokens. Output pricing includes reasoning tokens. The API supports low, medium and high reasoning effort, tool calling and JSON Schema output.

Editor's Summary

StepFun’s Step 5 Preview is now available as open weights in FP8 and BF16, with a 600B-parameter MoE design, 27B active parameters and 1M-token context aimed at agentic workflows. Its strongest cited independent result is 88.3% on AA-LCR, while its weakest cited result is 33.3% on Terminal-Bench 4.0, and the model requires data-center GPU deployments. The API is priced at $1.00 per 1M input tokens and $2.70 per 1M output tokens.