[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-stepfun-step-5-preview-pareto-frontier":3,"topics-all":32,"news-related-c05a1cb5-8f78-4729-812b-2e879284a51e":51},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":24,"news_slug":25,"published_at":26,"created_at":27,"modified_at":28,"is_published":29,"publish_type":30,"image_url":14,"view_count":31},"c05a1cb5-8f78-4729-812b-2e879284a51e","阶跃开源 600B MoE:输出低至每百万 2.7 美元","阶跃星辰 10 月 15 日开源 Step 5 Preview:600B MoE、27B 激活、1M 上下文,输出低至 2.7 美元\u002F百万 token,仅 Kimi K3 的 18%。表格与 Kimi K3、GLM 5.3 在编程\u002FAgent 上拉平,距 GPT-6 Astra、Claude Opus 5 一档。","阶跃星辰选在美东时间凌晨 1:30 左右,也就是北京时间 10 月 15 日下午,把 Step 5 Preview 的完整权重发到了 Hugging Face 和 ModelScope。BF16、FP8(手动审批)、NVFP4 三档 checkpoint 同时上架,StepFun Model License 1.0 跟着模型卡放了出来。这是 9 月 20 号 X 上一条不到 200 字预告之后的承诺落地。\n\n## 规格:超稀疏 MoE、1M 上下文、视频输入\n\nStep 5 Preview 是一颗稀疏 MoE:600B 总参数,每个 token 激活 27B,激活比例 4.5%。上下文窗口做到 1M tokens,单次输出上限 64K,文本、图像和视频三种输入都原生支持。模型 id 是 step-5-preview,跑在 OpenAI 兼容的 Chat Completions 接口上,Messages API 也开着。工具调用、流式输出、JSON Mode、JSON Schema、prompt cache 这些标准动作没缺。推理强度可以在请求里设 low \u002F medium \u002F high,把模型当客户来谈而不是当接口来调用,是这个接口比较特别的一点。\n\n## 价格:这场发布会的支点\n\nStepFun 自己 API 上的价格是输入 cache miss $1.00\u002F百万 token、cache hit $0.05、输出 $2.70,推理 token 一并按输出计。和同档旗舰放在一张表里看:Moonshot 自己的 Kimi K3 输出 $15.00,Step 5 是它的 18%;DeepSeek V4.1 Flash 离峰 $0.60、Z.ai 的 GLM-5.3-Flash $0.50、StepFun 自己上一代 Step 3.7 Flash $1.15。放到 Pareto 图上,Step 5 Preview 的位置是把同档智能的输出成本砍到了原来的 1\u002F3 到 1\u002F5,而不是顶着分数榜往上爬。StepFun 给自己的 Artificial Analysis Intelligence Index 打了 44 分,是同一档智能里最便宜的那一颗。\n\n## Benchmark:看 footnote 才稳\n\nSelf-report 的 benchmark 表要看三个 footnote 才稳:Step 5 Preview 跑在自家 High 档,对手跑在 Max 档,所以分数差被天然压了一点;有 6 行(包括 StepCodeBench、FinStepBench 三个子项)是 StepFun 自己研发,虽然开源代码但还没人复现;HLE with tools 这一行 Step 5 和 GLM 5.3 跑的是文本子集,别人跑全量,不可直接比较。把这些 footnote 抹掉,看到的是:DeepSWE v1.1 67.7,跟 Kimi K3 67.5 和 GLM 5.3 66.9 一档,但比 GPT-6 Astra 的 74.1、Claude Opus 5 的 74.0 落后 6-7 分;Terminal-Bench v4 33.3 反而压住 Kimi K3 的 12.6,但距离 Astra 57.9 一整档;FrontierFinance 66.4,只输给 Opus 5,跟金融赛道的闭源顶端挨得最近;BrowseComp 88.7 排第三,MMMU-Pro 76.0 在多模态上也只是个中上。所以同档智能的判断站得住:与 Kimi K3、GLM 5.3 拉平,跟 Astra、Opus 5 拉开。\n\n## 三个长任务实验,才是 StepFun 想卖的产品形态\n\n第一个是 24 小时在 H100 上从零优化 MLA GPU kernel(batch=1, 64 head, 8192 tokens),最佳一次跑到 508 TFLOPS,Claude Opus 5 同条件是 493 TFLOPS,3% 的领先是 StepFun 自己跑出来的实验,口径上要打折。第二个是 24 小时对一个 Qwen3-30B-A3B 基模做自动后训练,目标是 AIME24,从前 53.3% 提到 60%,Opus 5 同样任务接近,但是用了更少的标注 token。第三个是不给专项优化的宝可梦红版挑战,到第 3082 回合打完三枚道馆徽章、推进约 1\u002F3 进度。这三个实验叠加着说明的是 StepFun 想卖的产品形态:一个能在闭环里跑一天的模型,自己检查、自己回滚,价格又低到可以真的跑一天。\n\n## License:开源的 Apache-style,商用分两档\n\nLicense 这块不能漏。StepFun Model License 1.0 看起来是 Apache-2.0 风格,但有两个数字门槛:产品月活过 1 亿或者月收入过 2000 万美元,必须署 \"Powered by StepFun AI\";设备厂商把它当内嵌功能卖,或者 maaS 服务商 12 个月累计收入过 2000 万美元,要单独拿商用 license。这个分级写得明确,所以大厂拿 BF16 自部署之前要先算自己属不属于第二档。\n\n## 所以呢:Pareto 前沿向\"够用\"那一层推\n\nPareto 前沿这条线近两个月被 GLM 5.3、Kimi K3、Qwen3.8 一起推到中端,价格跌得猛、分数也压得住;Step 5 Preview 是把这波推到 Step 5 这个具体型号上,把 600B 总参数、27B 激活的 MoE 放到 $2.70 输出这一档——基本上是把\"day-long agent\"做成主营业务来定价,而不是按 token 卖 API。Astra 和 Opus 5 在顶,Step 5 在把\"够用\"这一层往下拽。要试的团队现在 BF16 1214.9 GB 一台 B300 节点或两台 H200 节点就能跑起来,NVFP4 单机 394.2 GB,门槛比 2.8T 的 Kimi K3 低一档。","https:\u002F\u002Fwww.stepfun.com\u002Fstep-5-preview","5f7d17cd-f95b-4a76-be2e-db79144de285",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"a8002d98-9df1-4ab9-94d4-a7625af634c4","china-ai",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"7e89b5cc-57db-4f37-bc6d-28919a73931c","model-release",{"id":19,"name":20,"slug":20,"description":14,"color":14},"d11f0044-8aef-487c-bebe-89ce4683a4a3","moe",{"id":22,"name":23,"slug":23,"description":14,"color":14},"b9bd9039-fcdb-41a8-b85b-fc1587def2b9","open-source",[],"stepfun-step-5-preview-pareto-frontier","2026-10-15T05:00:00Z","2026-10-15T03:45:36.095857Z","2026-10-15T03:45:36.095868Z",true,"agent",142,[33,42],{"slug":34,"tag_slug":34,"title_zh":35,"title_en":36,"intro_zh":37,"intro_en":38,"id":39,"is_active":29,"created_at":40,"modified_at":41},"ai-for-science","AI for Science 2026：从 UniPert 到 GPT-Rosalind 的硬核进化","AI for Science 2026: from UniPert to GPT-Rosalind","生命科学、化学材料、物理世界模型——AI 正在从\"语言工具\"变成\"实验伙伴\"。本专题收录 AI 在三大科学方向的关键节点：UniPert 统一基因与化学扰动空间、GPT-Rosalind 端到端生命科学推理、达摩院 AI 智能体 28 小时找到 4 种超导新材料、Anthropic Claude Science 把工作台做成标准品。","From language tool to lab partner — AI is reshaping life sciences, chemistry\u002Fmaterials, and physical world models. This topic covers the key milestones: UniPert unifying genetic-chemical perturbation spaces, GPT-Rosalind's end-to-end life-sciences reasoning, DAMO's AI agent discovering 4 superconducting materials in 28 hours, and Anthropic's Claude Science workbench going mainstream.","988a4300-5fab-41c4-b5d8-63711a2dc757","2026-09-10T01:34:15.296649Z","2026-09-10T01:34:15.296663Z",{"slug":43,"tag_slug":43,"title_zh":44,"title_en":45,"intro_zh":46,"intro_en":47,"id":48,"is_active":29,"created_at":49,"modified_at":50},"h3-series","MiniMax H3 系列：从开源权重到 35 倍吞吐","MiniMax H3 Series: from open weights to 35x throughput","MiniMax H3 自 2026 年 8 月开源以来节奏密集：官方把生成、参考与编辑收回一个模型；ComfyUI 当天压进 RTX 3060；摩尔线程 3 小时完成国产 GPU 适配；fal 后训练版把吞吐拉到 35 倍；FastH3 蒸馏再砍推理成本。本专题持续追踪 H3 的发布—开源—蒸馏—部署全链路。","Since MiniMax open-sourced H3 in August 2026 the pace has been relentless: one unified omni-modal model, same-day ComfyUI support down to an RTX 3060, a 3-hour Day-0 port to Moore Threads GPUs, fal's post-trained H3 Max at 35x throughput, and FastH3 distillation cutting inference cost further. This topic tracks the full H3 chain — release, open weights, distillation, deployment.","83ef0daa-3c31-4cb3-86ed-e5ee58654d5f","2026-09-08T07:33:19.942193Z","2026-09-08T07:33:19.942209Z",{"items":52},[53,58,63,68,73,78],{"id":54,"title":55,"news_slug":56,"published_at":57},"804ab59a-a8d6-4b61-bf74-8f6f2bdae83c","智谱把 Flash 做成一件正经事:一次说清 GLM-5.3-Flash 的架构和 benchmark 真相","glm-5-3-flash-hybrid-attention-architecture","2026-08-27T08:00:00+00:00",{"id":59,"title":60,"news_slug":61,"published_at":62},"b44fb160-9027-44cd-b355-a66869004abc","Nemotron 3.5 Super VL 开源:120B 总参只激活 12B","nemotron-3-5-super-vl-open-120b-moe","2026-10-20T17:11:56+00:00",{"id":64,"title":65,"news_slug":66,"published_at":67},"bbfcbee4-8d92-4078-bb1e-eaf4d1c1e146","Beam 权重落地:501B MoE 三精度齐发","beam-1-0-weights-open-agent-stack","2026-10-18T21:10:09+00:00",{"id":69,"title":70,"news_slug":71,"published_at":72},"d529114c-bfd7-4efb-9af4-444d3065b397","Step 5 Preview 开源,2.7 美元\u002F百万输出","stepfun-step-5-preview-600b-moe","2026-10-17T21:12:04+00:00",{"id":74,"title":75,"news_slug":76,"published_at":77},"cee60a45-f212-4a78-83dd-0226c07b3460","腾讯开源记忆写入模型 Hy-Memory:30B MoE 只激活 3B,四框架平均分超 Opus 5","tencent-hy-memory-30b-memory-writer","2026-10-17T15:14:21+00:00",{"id":79,"title":80,"news_slug":81,"published_at":82},"ac666c4e-4756-43d8-840a-4eacbf88f804","Meta 开源 816B MoE 模型 Muse Spark Open,跑在两台 Mac Studio 上","meta-muse-spark-open-816b-mac-studio","2026-10-17T03:00:00+00:00"]