[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-stepfun-step-5-preview-600b-moe":3,"topics-all":41,"news-related-d529114c-bfd7-4efb-9af4-444d3065b397":60},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":27,"news_slug":34,"published_at":35,"created_at":36,"modified_at":37,"is_published":38,"publish_type":39,"image_url":14,"view_count":40},"d529114c-bfd7-4efb-9af4-444d3065b397","Step 5 Preview 开源,2.7 美元\u002F百万输出","StepFun 推出 600B\u002F27B MoE 模型 Step 5 Preview,10 月 15 日三档权重开源,API 价 1\u002F2.7 美元每百万 token,DeepSWE 上贴近 GLM-5.3 与 Kimi K3。","阶跃星辰(StepFun)推出「600B 总参 \u002F 27B 激活」的稀疏 MoE 模型 Step 5 Preview,主打「同档智能、更低单位成本」。\n\n## 模型与价格\n\n1M token 上下文、64k 最大输出、三档推理强度,文字与图片输入(API 已扩到视频),工具调用与 JSON Schema 都有。架构 92 层窄深 Transformer,门控滑动窗口与全注意力 3:1 交错,MTP-3 推测解码提速约 30%。每 token 激活 27B 参数,FP8 推理最少 4 张 B300 或 8 张 H200。StepFun API 输入 1.00 美元\u002F百万 token(缓存 0.05),输出 2.70 美元\u002F百万 token(含推理)。横向:GLM-5.3-Flash 0.50、DeepSeek V4.1 Flash 0.60、Step 3.7 Flash 1.15、Kimi K3 15.00。Step 5 的旗舰对比组是 Kimi K3、GLM 5.3、GPT-6 Astra 与 Claude Opus 5——「同档智能」叙事下,2.70 把曲线往外推了一格。\n\n## 基准与打折处\n\nDeepSWE v1.1 上 67.7,GLM-5.3 是 66.9、Kimi K3 是 67.5、Claude Opus 5 是 74.0、GPT-6 Astra 是 74.1——三家中段几乎平手,落后美方旗舰 6-7 个百分点。Terminal-Bench 4.0 上 33.3,自建 StepCodeBench 上 49.0,优于 GLM-5.3 与 Kimi K3;但真实文档 GDP.pdf 只有 14.8,Astra 是 31.0。StepFun 自己写明:Step 5 跑 High、对手 Max(工作量不对等);六行带刀号的指标含 StepCodeBench 与 FinStepBench(都是 StepFun 自家设计);HLE-with-tools 那行 Step 5 与 GLM 5.3 跑文本子集、对手跑全集,StepFun 自己标「不直接可比」。\n\n## 开源落地\n\n10 月 15 日三档权重上 Hugging Face 与 ModelScope:BF16(1,214.9 GB)与 NVFP4(394.2 GB)直接下载,FP8(650.32 GB)需手动审批。许可证是 StepFun Model License 1.0,Apache-2.0 风格加商业条款(月活 1 亿以上或月营收 2,000 万美元以上需打 Powered by StepFun AI 标识;设备厂或 MaaS 年营收超 2,000 万美元需单独谈商业许可)。同日 DeepInfra 在 OpenRouter 开第二家供应商,价格一致。StepFun 公开的三个 24 小时级运行(H100 优化 MLA 核跑到 508 TFLOPS、AIME24 自动后训练、3,000 多回合 Pokémon Red)指向一件事:Step 5 是「能在合理预算下跑一整天」的代理循环,不是一次性 Q&A。\n\n## 所以呢\n\nStep 5 Preview 是「前沿智能价格曲线又被往下推一截」的样本,不是「又一个大模型」。对开源阵营是「价格战还没打完」的提醒;对美方旗舰是「继续压低单位 token 成本」的市场压力;对自托管派是 1.2 TB BF16 或 394 GB NVFP4 摆在面前的现实选项——门槛高,但许可证没把商业路堵死。技术报告暂未发布,这是这次发布最值得记一笔的留白。","https:\u002F\u002Fwww.marktechpost.com\u002F2026\u002F10\u002F16\u002Fstepfun-releases-step-5-preview-open-weights-600b-moe-27b-active-1m-context\u002F","5f7d17cd-f95b-4a76-be2e-db79144de285",[11,15,18,21,24],{"id":12,"name":13,"slug":13,"description":14,"color":14},"120fa59a-ff6f-4537-9bf5-f818df636a0e","benchmark",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"01598627-1ea6-4b27-a5d8-874971571a71","llm",{"id":19,"name":20,"slug":20,"description":14,"color":14},"7e89b5cc-57db-4f37-bc6d-28919a73931c","model-release",{"id":22,"name":23,"slug":23,"description":14,"color":14},"d11f0044-8aef-487c-bebe-89ce4683a4a3","moe",{"id":25,"name":26,"slug":26,"description":14,"color":14},"b9bd9039-fcdb-41a8-b85b-fc1587def2b9","open-source",[28],{"id":29,"lang":30,"title":31,"summary":32,"content":33},"d484ca04-ef11-4413-bffc-0817bcea763e","en","StepFun Releases Step 5 Preview: 600B MoE, Open Weights","StepFun ships Step 5 Preview, a 600B\u002F27B MoE. Three open-weight checkpoints on Oct 15 at $1\u002F$2.70 per million, with DeepSWE tracking GLM-5.3 and Kimi K3.","StepFun released Step 5 Preview, a sparse MoE model with 600B total parameters and 27B active per token, framed as \"comparable intelligence at a lower unit cost.\"\n\n## Model and pricing\n\n1M-token context, 64k max output, three reasoning-effort levels, text and image input (API extended to video), tool calling and JSON Schema supported. The architecture is a 92-layer narrow-deep Transformer, with gated sliding-window attention interleaved with full attention at a 3:1 ratio, and an MTP-3 speculative decoding head that accelerates end-to-end inference by about 30 percent. Each token activates 27B parameters, and FP8 inference requires a minimum of 4 B300 or 8 H200 GPUs. StepFun's API charges $1.00 per million input tokens (cache hit $0.05) and $2.70 per million output tokens (including reasoning tokens). Horizontally: GLM-5.3-Flash $0.50, DeepSeek V4.1 Flash $0.60, Step 3.7 Flash $1.15, Kimi K3 $15.00. Step 5's flagship comparison set is Kimi K3, GLM 5.3, GPT-6 Astra, and Claude Opus 5 — within the \"comparable intelligence\" framing, $2.70 pushes the cost curve outward by a notch.\n\n## Benchmarks and caveats\n\nOn DeepSWE v1.1, Step 5 scores 67.7, with GLM-5.3 at 66.9, Kimi K3 at 67.5, Claude Opus 5 at 74.0, and GPT-6 Astra at 74.1 — the three Chinese\u002Fopen-source mid-tier models are nearly tied, trailing US flagships by 6 to 7 points. On Terminal-Bench 4.0, Step 5 lands at 33.3; on its in-house StepCodeBench, 49.0, ahead of GLM-5.3 and Kimi K3; but on real-document GDP.pdf, only 14.8, versus Astra's 31.0. StepFun itself states: Step 5 was run at High effort while rivals ran at Max (workload mismatch); six dagger-marked benchmarks include StepCodeBench and FinStepBench, both StepFun's own designs; in the HLE-with-tools row, Step 5 and GLM 5.3 are run on a text-only subset while rivals are run on the full set, and StepFun itself labels the result \"not directly comparable.\"\n\n## Open weights and license\n\nOn October 15, three checkpoints landed on Hugging Face and ModelScope: BF16 (1,214.9 GB) and NVFP4 (394.2 GB) are downloadable directly; FP8 (650.32 GB) requires manual approval. The license is StepFun Model License 1.0, Apache-2.0 style with commercial add-ons (products above 100 million monthly active users or $20 million monthly revenue must display the Powered by StepFun AI attribution; device makers and MaaS businesses above $20 million in trailing-twelve-month revenue need a separate commercial license). The same day, DeepInfra became a second OpenRouter provider at the same price. StepFun published three 24-hour runs (H100 MLA kernel optimization reaching 508 TFLOPS, AIME24 automated post-training, and 3,000+ turns of Pokémon Red) — together they point to one thing: Step 5 is an \"affordable all-day agent loop,\" not a single-shot Q&A model.\n\n## What it means\n\nStep 5 Preview is \"the frontier-intelligence price curve being pushed down another notch,\" not \"yet another big model.\" For the open-source Chinese camp it's a reminder that the price war is not over; for US flagships it's market pressure to keep compressing per-token cost; for the self-host crowd it's 1.2 TB BF16 or 394 GB NVFP4 sitting in front of them as a real option — the threshold is high, but the license does not block the commercial path. No technical report has been published, and that is the most notable gap in this release.","stepfun-step-5-preview-600b-moe","2026-10-17T21:12:04Z","2026-10-18T05:12:12.593180Z","2026-10-18T05:12:12.593194Z",true,"agent",249,[42,51],{"slug":43,"tag_slug":43,"title_zh":44,"title_en":45,"intro_zh":46,"intro_en":47,"id":48,"is_active":38,"created_at":49,"modified_at":50},"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":52,"tag_slug":52,"title_zh":53,"title_en":54,"intro_zh":55,"intro_en":56,"id":57,"is_active":38,"created_at":58,"modified_at":59},"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":61},[62,67,72,77,82,87],{"id":63,"title":64,"news_slug":65,"published_at":66},"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":68,"title":69,"news_slug":70,"published_at":71},"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":73,"title":74,"news_slug":75,"published_at":76},"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":78,"title":79,"news_slug":80,"published_at":81},"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",{"id":83,"title":84,"news_slug":85,"published_at":86},"1b6ec126-cba1-47da-870b-da19cc523763","Meta 开源 816B MoE 模型 Muse Spark Open:1M 上下文,两台 Mac Studio 跑得动","meta-muse-spark-open-816b-moe-1m-mac-studio","2026-10-15T11:00:00+00:00",{"id":88,"title":89,"news_slug":90,"published_at":91},"5a5efb41-7037-46cb-b93e-6842f9ff2d62","UltraBERT:91.0 GLUE+47.7 BEIR,BERT 配方重审后 encoder 双榜登顶","ultrabert-encoder-dual-lm-heads","2026-10-14T17:10:59+00:00"]