[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"$fLSrszouJHmRxhSb4aDBTFFFkE2OQXKy9W-7hPGkJy9E":3,"$fWNsfe5LyMk_QM7fXcoufrWVu1aa3Gn4kGdkU0eJxsuI":82},{"id":4,"name":5,"code":6,"version":5,"logo":7,"url":8,"vendorId":9,"vendorName":10,"vendorCode":11,"vendorCountry":12,"modelType":17,"openSource":17,"openLicense":38,"paramSize":39,"contextLength":40,"intro":41,"supportApi":17,"apiUrl":42,"pricingType":43,"freeQuota":44,"score":27,"isRecommend":45,"isHot":17,"click":46,"tags":47,"content":48,"pricings":49,"scores":50},44,"Step 5 Preview","step-5-preview","\u002Flogos\u002Fstepfun.svg","https:\u002F\u002Fwww.stepfun.com\u002Fstep-5-preview",33,"阶跃星辰","stepfun",[13,18,22,26,30,34],{"id":14,"name":15,"color":16,"type":17},4,"开源","#909399",1,{"id":19,"name":20,"color":21,"type":17},5,"支持API","#00B4A0",{"id":23,"name":24,"color":25,"type":17},8,"国产自研","#F56C6C",{"id":27,"name":28,"color":29,"type":17},9,"推理强化","#E65100",{"id":31,"name":32,"color":33,"type":17},10,"长上下文","#3B55D9",{"id":35,"name":36,"color":37,"type":17},11,"多模态理解","#9C27B0",null,"600B（MoE，激活 27B）",1000000,"阶跃星辰旗舰基座模型：600B 稀疏 MoE、每 token 激活 27B，支持 100 万 token 上下文，权重已开源，综合智能指数 44 分位列全球开源前三","https:\u002F\u002Fplatform.stepfun.com",2,"权重开源免费下载，可自部署",0,39,[],"\u003Cp>\u003Cstrong>Step 5 Preview\u003C\u002Fstrong> 是阶跃星辰于 2026 年 9 月 20 日发布的新一代旗舰基座模型，官方定位为「面向真实世界 Agentic 任务」的主力工作模型，重点面向 AI 编程、软件工程、专业知识工作与金融等场景。\u003C\u002Fp>\u003Cp>模型采用稀疏 MoE 架构，总参数量 600B，每 token 仅激活约 27B 参数，原生支持文本与视觉输入，并支持 100 万 token 上下文窗口。技术博客披露其采用 92 层窄深架构、Sparse GQA、训练稳定性优化、百万级可验证环境与长程任务强化学习，并将上下文压缩纳入端到端策略训练，让 Agent 能整理任务状态后继续执行。\u003C\u002Fp>\u003Ch3>关键表现\u003C\u002Fh3>\u003Cul>\u003Cli>Artificial Analysis 综合智能指数 44 分，位列全球开源模型前三\u003C\u002Fli>\u003Cli>官方称在 DeepSWE v1.1、ProgramBench、FrontierFinance 等 8 项评测中有 6 项位居开源第一\u003C\u002Fli>\u003Cli>单任务成本约为同级闭源旗舰模型的八分之一\u003C\u002Fli>\u003Cli>生成速度约每秒 100 个 token，高于平均水平（约 65）\u003C\u002Fli>\u003C\u002Ful>\u003Ch3>开源与部署\u003C\u002Fh3>\u003Cp>2026 年 10 月 15 日，阶跃星辰正式开放 BF16 与 FP8 两个版本的模型权重，同步发布技术博客与配套开源开发工具，权重已上线 Hugging Face 与 ModelScope。开发者可通过 vLLM 或 SGLang 部署兼容 OpenAI 接口的 API 服务，也可使用 Transformers 进行离线推理。华为昇腾、壁仞、沐曦、摩尔线程、海光 DCU 等在开源当天完成适配。\u003C\u002Fp>\u003Ch3>生态落地\u003C\u002Fh3>\u003Cp>该模型已接入腾讯 WorkBuddy 与字节 TRAE；2026 年 10 月 10 日曾登顶 OpenRouter Trending 榜，并位列日榜全球第二。\u003C\u002Fp>\u003Ch3>使用建议\u003C\u002Fh3>\u003Cp>适合长上下文、多轮工具调用与持续执行的 Agentic 场景，以及软件工程与金融分析类任务。600B 总参数意味着自部署门槛较高，轻量场景建议直接使用云端 API；模型仍为 Preview 版本，能力与接口可能调整，生产环境接入前建议先做回归测试。\u003C\u002Fp>",[],[51,55,59,64,68,72,77],{"metricId":17,"metricName":52,"metricCode":53,"maxScore":31,"score":54},"中文理解","chinese",8.8,{"metricId":43,"metricName":56,"metricCode":57,"maxScore":31,"score":58},"文本写作","writing",8.5,{"metricId":60,"metricName":61,"metricCode":62,"maxScore":31,"score":63},3,"逻辑推理","reasoning",9.2,{"metricId":14,"metricName":65,"metricCode":66,"maxScore":31,"score":67},"代码能力","coding",9.4,{"metricId":19,"metricName":69,"metricCode":70,"maxScore":31,"score":71},"多模态","multimodal",8.6,{"metricId":73,"metricName":74,"metricCode":75,"maxScore":31,"score":76},6,"长文本","long-context",9.5,{"metricId":78,"metricName":79,"metricCode":80,"maxScore":31,"score":81},7,"性价比","cost-performance",9.3,[83,85,86,90,92,93,94,97],{"id":19,"name":20,"color":21,"type":17,"toolCount":84},35,{"id":31,"name":32,"color":33,"type":17,"toolCount":9},{"id":43,"name":87,"color":88,"type":17,"toolCount":89},"免费增值","#409EFF",21,{"id":14,"name":15,"color":16,"type":17,"toolCount":91},18,{"id":35,"name":36,"color":37,"type":17,"toolCount":31},{"id":27,"name":28,"color":29,"type":17,"toolCount":31},{"id":60,"name":95,"color":96,"type":17,"toolCount":78},"付费使用","#E6A23C",{"id":23,"name":24,"color":25,"type":17,"toolCount":78}]