[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-meituan-longcat-2-5-agent-rl":3,"topics-all":38,"news-related-179d83d0-4df3-4a0f-ac3b-aa31043c91f2":57},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":24,"news_slug":31,"published_at":32,"created_at":33,"modified_at":34,"is_published":35,"publish_type":36,"image_url":14,"view_count":37},"179d83d0-4df3-4a0f-ac3b-aa31043c91f2","美团没加参数:LongCat-2.5终端基准3.0变38.5","美团 11 月 2 日上线 LongCat-2.5:1.6T 参数、约 48B 激活,规模同 2.0,全押长程智能体 RL,Terminal Bench v4.0 从 3.0 跳到 38.5,BrowseComp 91.9;视觉 token 进主干,W4A4 跑国产加速器。未宣布开放权重,开源线仍停在 2.0。","11 月 1 日,美团 LongCat 团队发布 LongCat-2.5 并放出完整技术报告,次日模型在 LongCat API 平台可用([官方发布页](https:\u002F\u002Fwww.longcatai.org\u002Fnews\u002Flongcat-2-5.html))。规格表几乎没动:总参数维持 1.6T,每 token 激活约 48B,1M 上下文、128K 输出,和 9 月的 Preview 一致。变的是训练目标——2.0 学会写代码,2.5 要「完成项目」:跨代码库、终端、浏览器、表格和设计画布,而且它看得见。\n\n## 参数不动,分数跳变\n\n官方对比表里最说明问题的是这几个数字:Terminal Bench v2.1 从 70.8 到 88.4,v4.0 从 3.0 到 38.5,DeepSWE v1.1 从 9.7 到 64.6,PaperBench 从 45.2 到 82.1,BrowseComp 从 79.9 到 91.9。同一副骨架,最难的多步任务分数翻了几倍——官方把话说透:起作用的是长程智能体 RL,不是预训练规模。报告也没回避短板:对比集里 Terminal Bench v4.0 最优是 Claude Opus 5 的 53.9,高出约 15 分;DeepSWE v1.1 与 Terminal Bench v2.1 头名是 DeepSeek V4.1-Flash;纯视觉基准中游(VideoMME 85.2 对 Kimi K3 的 90.0)。官方结论很直白:广度有竞争力、带视觉编码特色,但最难的工程任务尚未登顶。\n\n## 视觉进主干\n\n「原生多模态」不是外挂视觉编码器:图像表示为多级离散 token(RVQ 量化原生分辨率 SAE 特征,LongCat-Next 路线),文本与视觉共享一套 token 流和训练目标。视觉 token 离线预计算并缓存,视觉编码器干脆不进训练图。在约 3B 激活参数的探针模型上,t-SNE 显示视觉与文本嵌入交织而非分簇。落地形态叫「视觉编码」:读截图、图表、UI 或 3D 视口,在同一个推理循环里写代码。\n\n## 国产加速器与 W4A4\n\n基础设施侧,长上下文训练和 W4A4 推理都在国产加速器上优化。1.6T MoE 的推理把专家权重和激活压到 4bit,靠 FlatQuant 的可学习等价变换恢复精度,官方称全部评测任务与未量化模型持平;线上压测对 W8A8,prefill 吞吐 +43.3%,decode +42.77%,TTFT 降低 15.91%。训练侧 FSDP 配置在长序列 SFT 上比 DDP 基线每卡吞吐高 86%。\n\n## 开源停在 2.0\n\n最值得玩味的是发布方式:未宣布开放权重。这打破了 LongCat 家族的默认模式——Flash-Chat、Flash-Thinking、2.0(MIT)、Next、Video-Avatar 都放过权重。对自托管用户,2.0 仍是可用的最新开放权重;2.5 是托管服务,带相应的延迟、数据驻留与计价依赖。第三方报道([TechNode 经 AI Weekly 转述](https:\u002F\u002Faiweekly.co\u002Falerts\u002Fmeituan-ships-longcat-25-a-16t-param-multimodal-agent-model))确认 11 月 2 日上线与国产芯片训练口径;权重放不放,等下几次发布给答案。\n\n所以呢:当参数不再增长、Terminal Bench v4.0 还能从 3.0 跳到 38.5,「堆预训练 token」的叙事就该让位了——下一轮比拼的是谁能在真实环境里把 RL 跑通。你是继续等开放权重,还是先换个 base URL 试试 API?","https:\u002F\u002Fwww.longcatai.org\u002Fnews\u002Flongcat-2-5.html","a35708a6-b30c-4d13-96b5-7931272da6a8",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"6ad31a14-c0da-42df-81fd-564281f768db","agentic-ai",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",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"c138c058-98f6-4bdc-87f8-92edce268f55","en","LongCat-2.5: Same 1.6T Params, Terminal Bench 3.0 to 38.5","LongCat-2.5 keeps 1.6T\u002F48B active but lifts Terminal Bench v4.0 from 3.0 to 38.5 via long-horizon RL. API-only; open weights stay at 2.0.","On November 1, Meituan's LongCat team released LongCat-2.5 with a full technical report; the model went live on the LongCat API platform the next day ([official release page](https:\u002F\u002Fwww.longcatai.org\u002Fnews\u002Flongcat-2-5.html)). The spec sheet barely moved: 1.6T total parameters, roughly 48B active per token, 1M context, 128K output — the same envelope as the September preview. What changed is the training target: 2.0 was built to write code, 2.5 is built to finish projects — across a codebase, a terminal, a browser, a spreadsheet, or a design canvas. And it can see.\n\n## Same parameters, different scores\n\nThe official comparison against LongCat-2.0 tells the story: Terminal Bench v2.1 70.8 to 88.4, v4.0 3.0 to 38.5, DeepSWE v1.1 9.7 to 64.6, PaperBench 45.2 to 82.1, BrowseComp 79.9 to 91.9. Same skeleton, multi-step scores multiplied several times over — and the team is explicit in the report that long-horizon agentic RL, not pretraining scale, did the work. To its credit, the report does not hide the gaps: in the comparison set Claude Opus 5 tops Terminal Bench v4.0 at 53.9, about 15 points ahead; DeepSeek V4.1-Flash leads DeepSWE v1.1 and Terminal Bench v2.1; pure-vision results sit mid-pack (VideoMME 85.2 against Kimi K3's 90.0). The official summary is blunt: competitive breadth with a visual-coding edge, not yet dominant on the hardest engineering tasks.\n\n## Vision in the backbone\n\n\"Native multimodal\" here means no bolted-on vision encoder: images become multi-level discrete tokens — RVQ over native-resolution SAE features, the LongCat-Next lineage — so text and vision share one token stream and one objective. Visual tokens are precomputed and cached offline, which means the vision encoder never enters the training graph at all. On a roughly 3B-active probe model, t-SNE shows visual and text embeddings interwoven rather than clustered apart. The shipped form is \"visual coding\": read a screenshot, chart, UI, or 3D viewport inside the same reasoning loop that writes the code.\n\n## W4A4 on domestic accelerators\n\nOn the infrastructure side, long-context training and W4A4 inference are both tuned for domestic Chinese accelerators. Inference quantizes expert weights and activations to 4 bits, with FlatQuant-style learnable equivalent transformations recovering accuracy; the company says the quantized model matches the unquantized one on every evaluated downstream task. Against W8A8 in online stress tests: prefill throughput +43.3%, decode +42.77%, TTFT down 15.91%. On the training side, the FSDP configuration delivers 86% higher per-card throughput than a DDP baseline on long-sequence SFT.\n\n## Open weights stop at 2.0\n\nThe most telling detail is distribution: no open-weight release has been announced. That breaks the family default — Flash-Chat, Flash-Thinking, LongCat-2.0 (MIT), Next, and Video-Avatar all shipped weights. For self-hosters, the practical reality is that 2.0 remains the newest usable open weight, and 2.5 has to be treated as a hosted service with the usual latency, data-residency, and pricing dependencies. Third-party coverage ([TechNode via AI Weekly](https:\u002F\u002Faiweekly.co\u002Falerts\u002Fmeituan-ships-longcat-25-a-16t-param-multimodal-agent-model)) confirms the November 2 availability and the domestic-chip training angle, but whether weights follow is a question for the next releases.\n\nSo: when parameters stop growing and Terminal Bench v4.0 still jumps from 3.0 to 38.5, the \"more pretraining tokens\" story yields to whoever can actually run RL in real environments. Wait for weights, or swap a base URL and try the API?","meituan-longcat-2-5-agent-rl","2026-11-03T15:09:34Z","2026-11-03T15:09:40.084538Z","2026-11-03T15:09:40.084545Z",true,"agent",172,[39,48],{"slug":40,"tag_slug":40,"title_zh":41,"title_en":42,"intro_zh":43,"intro_en":44,"id":45,"is_active":35,"created_at":46,"modified_at":47},"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":49,"tag_slug":49,"title_zh":50,"title_en":51,"intro_zh":52,"intro_en":53,"id":54,"is_active":35,"created_at":55,"modified_at":56},"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":58},[59,64,69,73,78,83],{"id":60,"title":61,"news_slug":62,"published_at":63},"9753c773-a3c7-4329-9c09-f2037eb1c086","IQuest-Q1开源:320B MoE专攻CLI智能体","iquest-q1-320b-moe-agentic-cli","2026-10-22T09:01:29+00:00",{"id":65,"title":66,"news_slug":67,"published_at":68},"554f6552-6c09-401a-99e3-c9ce60a497ca","ERNIE 5.5 基座:280B激活14B,对话版还没来","baidu-ernie-5-5-flash-base-open-weights","2026-11-05T12:00:00+00:00",{"id":70,"title":71,"news_slug":72,"published_at":68},"54b66da0-9f30-4777-b7d5-c19df1ecf6d1","Inkling 2 Small：训练中绕过封锁被监控抓到","inkling-2-small-network-block-rl-monitoring",{"id":74,"title":75,"news_slug":76,"published_at":77},"32560a11-3d59-4397-8f01-e75f680a59db","月之暗面兑现承诺:K3.1 Pro 2.8T 权重开源上架","kimi-k3-1-pro-open-weights-huggingface","2026-11-04T23:07:35+00:00",{"id":79,"title":80,"news_slug":81,"published_at":82},"26337a90-e049-44cb-9738-16d10a1fe84b","微软4B混合智能体:OSWorld 66%,点击还是敲代码","microsoft-mage-vl-cowork-hybrid-computer-use","2026-11-03T07:33:53+00:00",{"id":84,"title":85,"news_slug":86,"published_at":87},"4d0566f5-ef08-4c5f-befc-09dde8010620","Le Chonk 开源兑现:三周RL让TB冲上41%","mistral-large-4-open-weights-terminal-bench","2026-10-29T17:11:35+00:00"]