[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-cohere-north-small-translate":3,"topics-all":38,"news-related-8e730a3d-439b-45cf-961d-f77cf01469fd":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},"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 开源翻译专用模型 North Small Translate:218B 总参\u002F25B 激活 MoE,覆盖 50 种语言。在自测 WMT26 评测得 83.60,超过 DeepL NextGen 的 81.37;长文档得分为 Google 翻译两倍,4bit 版两块 H100 即可部署。","Transformer 在 2017 年为翻译而生——Google 提出它,是为了替换自家的机器翻译架构。九年后,由这篇论文作者参与创立的 Cohere,绕回了起点:发布翻译专用大模型 North Small Translate,权重开放在 Hugging Face。\n\n## 发布了什么\n\nNorth-Small-Translate-1.0 是一个稀疏 MoE 解码器:总参数 218B,每个 token 只激活 25B。架构上共 128 个专家,每 token 选 8 个,外加共享专家;注意力层以 3:1 比例交错滑动窗口注意力(窗口 4096,带 RoPE)与全局注意力(无位置编码),沿用 Command A 的设计;路由器对专家 logits 做 sigmoid 激活后在选中的 top-k 上归一化。\n\n它覆盖 50 种语言(32 种高资源 + 18 种其他),上下文 16K 输入 \u002F 16K 输出,纯文本进出。官方提供三档量化:BF16 需 4×B200 或 8×H100,FP8 需 2×B200 或 4×H100,4-bit NVFP4 版本 1×B200 或 2×H100 即可跑。许可证为 CC BY-NC 4.0——研究开放,商用走 RWS 的 Language Weaver 平台。\n\n## 跑分:专科生对阵通才\n\n在 Cohere 自己运行的 WMT26 全语言评测(以 GPT-5.6-Sol 为评审)上,North Small Translate 得分 83.60,领先 Qwen 3.5 397B A17B(81.56)、DeepL NextGen(81.37)、Gemma 4 31B(79.46)、GLM 5.2 FP8(76.50)与 Google 翻译(68.20);支持自查自纠的「Agentic」版本进一步到 84.36。\n\n长文档是差距最大的地方:单次调用翻译两章书,它得 48.9 分,Google 翻译 21.3、Gemma 4 31B 19.4——超过通才模型的两倍。\n\n## 效率账\n\n同等硬件与并发下,它的输出吞吐达到每秒 112 token,Gemma 4 31B 为 81(高并发下 39 对 30),多出约 30-38%。商用配置平均每任务只耗 661 token、$0.000676;Cohere 给出的对照组 Gemini 3.1 Pro Preview(high)每任务 $0.038928,贵 5762%。\n\n## 冷水先泼\n\n以上数字全部出自 **Cohere 自测**。MarkTechPost 提醒,在独立 WMT26 结果出现前应视为厂商自报;基准聚合站 BenchLM 同样只作展示、不纳入加权排名。市场反应也偏冷清:HF 模型卡显示近一个月下载量约 40 次。\n\n但信号是真的:继 Command A+、North Mini Code 之后,这是 Cohere 四个月内第三批开放权重发布,沿用同一套 218B\u002F25B 骨架,这次打的是 DeepL 的主场——主权 AI 场景(本地化、政务、受监管行业)要求数据不出境,过去只能在闭源 API 与通用模型兼职之间二选一,现在多了一个在厂商自测榜单上压过 DeepL、且两块 H100 装得下的选项。\n\n九年前 Transformer 是为了让机器翻译而设计的;后来通用化吞掉了一切专科。反攻,从一张装进两块 H100 的模型卡开始。\n\n参考:[Cohere 官方博客](https:\u002F\u002Fcohere.com\u002Fblog\u002Fnorth-small-translate) · HF 模型卡(CohereLabs\u002FNorth-Small-Translate-1.0) · MarkTechPost 报道","https:\u002F\u002Fcohere.com\u002Fblog\u002Fnorth-small-translate","df9f8204-8e8d-4fce-8526-3c6fe8e6ae56",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"01598627-1ea6-4b27-a5d8-874971571a71","llm",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},"b1ab7011-66b0-4ddf-9b00-47edb4d4b45c","en","Cohere Open-Sources 218B Translation MoE, Tops DeepL on WMT26","Cohere Labs open-sources North Small Translate: a 218B\u002F25B-active MoE for 50 languages, 83.60 on its own WMT26 eval vs DeepL's 81.37.","The Transformer was born in 2017 for translation — Google proposed it to replace its machine-translation stack. Nine years later, Cohere, a company founded by authors of that paper, has circled back to the starting point: North Small Translate, a dedicated translation LLM with open weights on Hugging Face.\n\n## What shipped\n\nNorth-Small-Translate-1.0 is a sparse Mixture-of-Experts decoder: 218B total parameters, only 25B active per token. The architecture has 128 experts, with 8 selected per token plus shared experts. Attention layers interleave sliding-window attention (window 4096, with RoPE) and global attention (no positional embeddings) at a 3:1 ratio, as first introduced in Command A; the router applies sigmoid activation over the expert logits and normalizes over the selected top-k.\n\nIt covers 50 languages (32 high-resource plus 18 additional), with 16K input \u002F 16K output context, text in and text out. Three quantizations ship: BF16 needs 4x B200 or 8x H100, FP8 needs 2x B200 or 4x H100, and the 4-bit NVFP4 build fits on 1x B200 or 2x H100. The license is CC BY-NC 4.0 — open for research, with commercial use routed through RWS's Language Weaver platform.\n\n## Benchmarks: specialist vs generalist\n\nOn the WMT26 all-languages evaluation Cohere ran itself (with GPT-5.6-Sol as judge), North Small Translate scores 83.60 — ahead of Qwen 3.5 397B A17B (81.56), DeepL NextGen (81.37), Gemma 4 31B (79.46), GLM 5.2 FP8 (76.50) and Google Translate (68.20). The \"Agentic\" variant, which finds and fixes its own translation errors, reaches 84.36.\n\nLong documents are where the gap widens most: translating two book chapters in a single call, it scores 48.9 versus Google Translate's 21.3 and Gemma 4 31B's 19.4 — more than double the generalists.\n\n## The efficiency math\n\nUnder identical hardware and concurrency, output throughput hits 112 tokens per second against Gemma 4 31B's 81 (39 vs 30 at high concurrency) — roughly 30-38% more. The commercial configuration averages just 661 tokens per task at $0.000676; the comparison point Cohere gives, Gemini 3.1 Pro Preview (high), costs $0.038928 per task — 5,762% more.\n\n## A bucket of cold water\n\nEvery number above is **Cohere-run**. MarkTechPost advises treating the scores as vendor-reported until independent WMT26 results appear, and benchmark aggregator BenchLM likewise keeps the model out of weighted rankings, display-only. Market response has been lukewarm: the HF model card shows roughly 40 downloads in the past month.\n\nBut the signal is real: after Command A+ and North Mini Code, this is Cohere's third open-weight release in four months on the same 218B\u002F25B footprint — and this time it is aimed squarely at DeepL's home turf. For sovereign-AI deployments (localization, government, regulated industries) where data cannot leave the premises, the choice used to be between a closed API and a generalist model moonlighting as a translator. Now there is a third option that beats DeepL on the vendor's own benchmark and fits on two H100s.\n\nNine years ago the Transformer was built to make machines translate; then generalization swallowed every specialist. The counterattack starts with a model card that fits on two H100s.\n\nReferences: [Cohere blog](https:\u002F\u002Fcohere.com\u002Fblog\u002Fnorth-small-translate) · HF model card (CohereLabs\u002FNorth-Small-Translate-1.0) · MarkTechPost coverage","cohere-north-small-translate","2026-09-11T19:07:20Z","2026-09-11T19:07:22.810079Z","2026-09-11T19:07:22.810093Z",true,"agent",24,[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},"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":65,"title":66,"news_slug":67,"published_at":68},"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",{"id":70,"title":71,"news_slug":72,"published_at":73},"3d36921f-3b84-4663-97a0-fee7d4eff795","汤森路透开源 Thomson-1.0-Small:持续学习改造 Qwen,3B 激活的 35B MoE","thomson-1-0-small-continual-learning","2026-08-28T19:10:00+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"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":80,"title":81,"news_slug":82,"published_at":83},"7958a2f1-028c-4b4e-b134-0d5de9afc1c1","Motif 3 收官:韩国 314B MoE 改用 MIT 许可,从零起步架构首次面向商用","motif-3-mit-license-sovereign-ai","2026-08-24T00:00:00+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"491f4904-c854-4925-b3e3-e34b8afd5e50","KDA+MLA 混合栈下沉到 1.3B 激活:Ling-3.0-tiny 把 MoE 端侧化,INT4 跑出 115 tok\u002Fs","ling-3-tiny-kda-mla-edge-deployment","2026-08-18T00:00:00+00:00"]