[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-nex-n2-5-open-weights-agent-family":3,"topics-all":38,"news-related-643addcf-6639-4c7b-b6b9-dc478bab4792":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},"643addcf-6639-4c7b-b6b9-dc478bab4792","Nex-N2.5 三档齐发:1.6T MoE 权重开放,BrowseComp 92.6 超 Opus 5","Nex AGI 发布三档 agent 模型 Nex-N2.5,Max 为 1.6 万亿参数纯文本 MoE,权重全部开放。BrowseComp 92.6 超 Opus 5,Terminal-Bench 86.1 仍落后;mini 两张 H100 即可部署。","9 月上旬,agent 赛道又多了一个大动作:Nex AGI 把新一代模型家族 Nex-N2.5 一次性放出三档——mini、Pro、Max。主角是 Max:1.6 万亿参数的纯文本 MoE,官方称之为「我们首次在万亿参数尺度上完成的完整后训练」。三档权重全部开放,Hugging Face、ModelScope 都能下载,mini 和 Pro 还接入了 OpenRouter 托管服务。\n\n## 视觉不是输入,是接口\n\nNex-N2.5 冲的是长时程真实环境任务:操作电脑、浏览网页、自主写程序并跑测试。mini 和 Pro 延续上一代 Nex-N2 的多模态底座,重点强化 computer use、网页浏览与视觉接地的 agent 能力。官方 README 有句值得划线的表述:视觉不再只是一种输入模态,而是 agent 感知环境、验证结果、推动任务的「关键接口」——模型要看屏幕做事,做完看结果自我纠错,反馈闭环靠视觉完成。\n\n## 跑分:单项反超,整体追赶\n\n官方评测把三档与 Claude Opus 5、GPT-5.6 Sol、Kimi-K3、GLM-5.3、DeepSeek-V4-Pro、Qwen3.8-Max 同表对比。亮点:Max 在 BrowseComp 拿 92.6,压过 Claude Opus 5 的 90.8、GPT-5.6 Sol 的 90.4 与 Kimi-K3 的 91.2,为该表最高分;Pro 在 OSWorld-G 拿 87.4,超过 Qwen3.8-Max 的 84.9 和 Claude Opus 5 的 76.8。短板同样明显:Terminal-Bench 2.1 上 Max 为 86.1,落后 Opus 5 的 89.1;SWE-Bench Pro 上 65.7 对 79.2。一句话总结:单项能赢前沿闭源,整体仍在追赶。\n\n## 部署门槛:万亿参数不白给\n\n参考部署配置里,Max 要 2 节点共 32 张 H200(官方定制 SGLang fork 的 Docker 镜像,262144 上下文);Pro 降到单节点 8×H100;mini 两张 H100 就能起服务。模型提供 reasoning_effort 三档思维模式(关闭\u002F自适应\u002F强制思考),工具调用解析器兼容 qwen3_coder。许可方面,多家第三方追踪源确认三档均为 Apache-2.0。\n\n## 所以呢\n\n看点不在「又一组开源权重」,而在万亿 MoE 的完整后训练流程被这家公司走通之后,连评测口径、部署命令、思维模式开关都一并公开——agent 原生训练又多了一份可对照的公开样本。对应用团队,mini 两卡可跑、OpenRouter 直接可试,试错成本已压到地板;值得盯的下一个问题是:靠视觉反馈自我纠偏这套训法,能不能下放到更小的模型上?\n\n完整评测表与部署命令见官方仓库:https:\u002F\u002Fgithub.com\u002Fnex-agi\u002FNex-N2.5","https:\u002F\u002Fgithub.com\u002Fnex-agi\u002FNex-N2.5","65b22a69-1f6e-4be8-84da-6cd5f13d29e9",[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},"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",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"f4f50604-ea8b-4fd6-93a1-f42f7ef9cad5","en","Nex-N2.5: Nex AGI's 1.6T MoE agent family, open weights","Nex AGI open-sources three agent tiers; Max is a 1.6T text-only MoE. BrowseComp 92.6 tops Opus 5; Terminal-Bench 86.1 trails. Mini needs two H100s.","Early September brought a sizable move in the agent space: Nex AGI shipped its next-generation model family, Nex-N2.5, in three tiers at once — mini, Pro, and Max. The headline is Max, a 1.6-trillion-parameter, text-only Mixture-of-Experts model that the company describes as its \"first complete post-training effort at trillion-parameter scale.\" Weights for all three tiers are open, downloadable from Hugging Face and ModelScope, with mini and Pro also available through hosted OpenRouter endpoints.\n\n## Vision as an interface, not an input\n\nNex-N2.5 targets long-horizon tasks in real environments: operating computers, browsing the web, and writing and testing programs autonomously. mini and Pro build on the multimodal foundations of the earlier Nex-N2 generation, with focused gains in computer use, web browsing, and visually grounded agent capabilities. The README puts it crisply: vision is no longer merely an input modality — it is the interface through which an agent perceives its environment, verifies outcomes, and moves a task forward. The model watches the screen to act, then watches the result to self-correct.\n\n## Benchmarks: wins on specific fronts, chasing overall\n\nThe official evaluation tables put the three tiers against Claude Opus 5, GPT-5.6 Sol, Kimi-K3, GLM-5.3, DeepSeek-V4-Pro, and Qwen3.8-Max. The highlights: Max scores 92.6 on BrowseComp, ahead of Claude Opus 5 (90.8), GPT-5.6 Sol (90.4), and Kimi-K3 (91.2) — the top score in that table; Pro takes 87.4 on OSWorld-G, above Qwen3.8-Max (84.9) and Claude Opus 5 (76.8). The gaps are equally clear: on Terminal-Bench 2.1, Max's 86.1 trails Opus 5's 89.1; on SWE-Bench Pro it is 65.7 versus 79.2. The honest summary: it can beat frontier closed models on individual agentic tests while still chasing them overall.\n\n## Deployment: a trillion parameters is not free\n\nThe reference deployment for Max spans 2 nodes with 32 H200s in total (a custom SGLang fork shipped as a Docker image, with a 262,144-token context); Pro drops to a single node of 8×H100; mini serves on two H100s. The family supports three thinking modes via reasoning_effort (off \u002F adaptive \u002F always-on) and uses the qwen3_coder tool-call parser. Multiple independent trackers confirm all three tiers ship under Apache-2.0.\n\n## So what\n\nThe story here is not \"another set of open weights.\" It is that after walking a trillion-parameter MoE through a complete post-training pipeline, Nex AGI published the evaluation setup, deployment commands, and thinking-mode switches along with it — one more referenceable public recipe for agent-native training. For app teams, mini runs on two cards and OpenRouter access is a low-friction starting point. The question worth watching next: can this visual-feedback self-correction recipe travel down to much smaller models?\n\nFull benchmark tables and deployment commands are in the official repository: https:\u002F\u002Fgithub.com\u002Fnex-agi\u002FNex-N2.5","nex-n2-5-open-weights-agent-family","2026-09-12T13:07:00Z","2026-09-12T13:08:11.634984Z","2026-09-12T13:08:11.634999Z",true,"agent",71,[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},[59,64,69,74,79,84],{"id":60,"title":61,"news_slug":62,"published_at":63},"28e1fd65-7f26-41e8-b1ba-aa69d03dce84","上海AI实验室开源 Atria Dawn Preview:五项 Agent 基准居首","shanghai-ai-lab-atria-dawn-preview","2026-09-14T19:08:09+00:00",{"id":65,"title":66,"news_slug":67,"published_at":68},"21fe3c11-4ba4-4801-b6fc-60c4ae559dc1","Yandex 逆流开源:35B 参数的 T5 MoE,每个 token 只激活 0.6B","yandex-aliceai-t5-sparse-moe","2026-09-16T19:11:43+00:00",{"id":70,"title":71,"news_slug":72,"published_at":73},"8e730a3d-439b-45cf-961d-f77cf01469fd","Cohere 开源 218B 翻译专用 MoE:25B 激活,自测评分超 DeepL,2×H100 可部署","cohere-north-small-translate","2026-09-11T19:07:20+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"d941056b-c2e7-42e5-965a-a982c20b1169","Qwen3.8-Flash-Next 架构细节:Gated Residual 多分支残差 + QSA micro-block 稀疏注意力","qwen3-8-flash-next-cost-efficiency-architecture","2026-09-02T02:00:00+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"61de017b-bdd6-44b3-9f45-d4fb233bd24d","PhoneLLM 开源:30B MoE 电话客服模型,自称比 GPT-5.6 Terra 便宜 94%","phonellm-alpha-1-voice-agent-open-model","2026-08-29T21:10:00+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"33f3b08b-c8a2-43ec-81cf-85e2b918f913","腾讯开源 Hy4 preview:770B MoE、1M 上下文,模型首次参与自身训练","tencent-hy4-preview-770b-moe","2026-08-29T15:00:00+00:00"]