StepFun 开源 Step 5 Preview：600B MoE、27B 激活、1M 上下文
StepFun Releases Step 5 Preview Open Weights: 600B MoE, 27B Active, 1M Context - MarkTechPost
StepFun 以开放权重形式发布 Step 5 Preview，FP8 与 BF16 检查点已上线 Hugging Face 和 ModelScope。该模型为稀疏 MoE 架构，总参数 600B、每 token 激活 27B，上下文最长 1M token，面向编码智能体、长文档分析与金融场景。自托管最低需 4 张 B300（FP8）或 8 张 H200，API 定价为每 1M token 输入 1.00 美元、输出 2.70 美元；Artificial Analysis 的 AA-LCR 长上下文推理得分为 88.3%，Terminal-Bench 4.0 为 33.3%。
推荐理由
精选AI 评分 76
原文列出开放权重、自托管硬件门槛与多项基准成绩，便于读者判断这一开源模型的部署成本与能力边界。
主题
正文
当前显示原文，中文全文译文尚未就绪。
Add as a preferredsource on Google
StepFun has released Step 5 Preview as open weights. FP8 and BF16 checkpoints are live on Hugging Face and ModelScope. Step 5 Preview is a sparse Mixture-of-Experts model from StepFun built for long-horizon agentic work. It has 600B total parameters, 27B active per token and up to 1M tokens of context. For developers, this is a new self-hostable option for coding agents, document analysis and finance workflows.
TL;DR
Size: 600B total, 27B active per token, up to 1M-token context.
Runs on: data-center GPUs only. FP8 needs 4x B300 or 8x H200. BF16 needs 8x B300 or 16x H200.
Performance: near open-weight peers GLM-5.3 and Kimi K3 on coding. Behind GPT-6 Astra and Claude Opus 5.
Best: 88.3% on AA-LCR long-context reasoning (Artificial Analysis).
Worst: 33.3% on Terminal-Bench 4.0, versus 41.9% for GLM-5.3.
What is Step 5 Preview?
Step 5 Preview is StepFun’s flagship model for agentic work. StepFun positions it for software engineering, professional knowledge work and finance. The model navigates large codebases, coordinates multi-file changes, runs code and iterates on results. It also analyzes long documents, charts, screenshots and video to produce structured reports.
The API launched September 20. Weights followed on October 14.
How does the architecture work?
Only 27B of the 600B parameters activate per token. StepFun’s tech team adds the internals:
Depth: 92 transformer blocks in a narrow, deep design.
Attention mix: gated sliding-window attention and full attention layers, interleaved 3:1.
Sparse GQA: full-attention layers pick 4,096 history tokens per query, scored in blocks of 8.
Speculative decoding: an MTP-3 head gives roughly 30% end-to-end speedup.
Training: FP8 mixed precision, with RL on million-level verifiable agentic environments.
How does Step 5 Preview perform on benchmarks?
StepFun ran its own evaluations at the High reasoning setting. Rivals in its chart ran at Max effort, so the comparison is not like-for-like.
Results reported on the Hugging Face card:
DeepSWE v1.1: 67.7 (SWE-agent harness, temperature 1.0, top_p 0.95)
Terminal-Bench 4.0: 33.3
Agents’ Last Exam (ALE-CLI): 29.5
Independent numbers come from Artificial Analysis via OpenRouter. The model scores 43.7 on the Intelligence Index. It hits 88.3% on AA-LCR and 46.5% on Humanity’s Last Exam. GDPval-AA lands at 54.3%.
Vals AI results show clear weak spots. SRE Bench is 0.4%. Harvey’s Legal Agent Benchmark is 9.2%.
On finance, StepFun’s chart shows 66.4% on FrontierFinance. That beats GLM-5.3 and Kimi K3, but trails Claude Opus 5 at 69.7%.
How does it compare with other open-weight models?
FeatureStep 5 PreviewGLM-5.3Kimi K3
Total parameters600B753B (HF file size)2.8T
Active parameters27BNot disclosed104B
Context window1M1M1M
Input modalitiesText, image (API adds video)Text onlyText, image
Weight formatsFP8, BF16BF16, FP8MXFP4 weights
LicenseStepFun Model License 1.0GLM-5.3 licenseKimi K3 License
Minimum listed self-host setup4x B300 (FP8)Not disclosedNot disclosed
DeepSWE v1.1 (StepFun-reported)67.766.967.5
Terminal-Bench 4.0 (StepFun-reported)33.341.912.6
FrontierFinance (StepFun-reported)66.464.162.6
API price per 1M tokens (in / out)$1.00 / $2.70Not disclosed$3.00 / $15.00
Benchmark rows come from StepFun’s comparison chart, as transcribed by CellCog and eesel.
What hardware does Step 5 Preview need to run?
This is a data-center model. The FP8 checkpoint alone is 650.32 GB. Listed configurations:
FP8 on vLLM or SGLang: 4 B300 GPUs or 8 H200 GPUs, each on 1 node.
BF16 on vLLM: 1 node with 8 B300 GPUs, or 2 nodes with 16 H200 GPUs.
Transformers: offline generation only. The full BF16 shard workflow is not yet validated end to end.
How much does the API cost?
The StepFun API charges $1.00 per 1M input tokens and $2.70 per 1M output tokens. Cache hits cost $0.10 per 1M. Output pricing includes reasoning tokens. The API supports low, medium and high reasoning effort, tool calling and JSON Schema output.
Key Takeaways
600B MoE, 27B active, 1M context, now open weights in FP8 and BF16.
Smallest listed self-host setup: 4x B300 GPUs with FP8.
Best independent score: 88.3% on AA-LCR long-context reasoning.
Weakest spot: 33.3% on Terminal-Bench 4.0.
API costs $1.00 in and $2.70 out per 1M tokens.
Michal Sutter+ postsBio ⮌
Michal Sutter is a data science professional with a Master of Science in Data Science from the University of Padova. With a solid foundation in statistical analysis, machine learning, and data engineering, Michal excels at transforming complex datasets into actionable insights.
Michal SutterMeet Grok Voice 3.0: SpaceXAI Voice Agent Model Ranks #1 on Artificial Analysis
Michal SutterZ.ai GLM-Legal Beats Claude Fable 5 and Kimi K3 on Harvey LAB
RELATED ARTICLESMORE FROM AUTHOR
DeepSeek Just Released V4.1-Flash-1017, Scores 77.0 on DeepSWE, Opens Search API
NVIDIA Releases Nemotron 3.5 Super VL, a 120B Open Multimodal Agent Model