[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-bilibili-index-translate-35b-moe":3,"topics-all":38,"news-related-9fa3427c-cf69-46c7-9720-cd3b646a155b":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},"9fa3427c-cf69-46c7-9720-cd3b646a155b","B站开源35B翻译模型:3B激活,150种语言","哔哩哔哩 Index LLM 团队开源 Index-Translate 翻译模型家族:旗舰 35B-A3B 为总参 35B、激活约 3B 的 MoE,基于 Qwen3.5,覆盖 150 种语言,Apache-2.0 放出权重。官方报告称其指令遵循领先同规模开源模型,低资源语向 off-target 率 2.4%。","翻译这个被认定\"已经被解决\"的老任务,最近又热闹了起来。腾讯混元、Cohere 相继开源翻译专用模型之后,哔哩哔哩 Index LLM 团队在 9 月 30 日交出了自己的答卷:Index-Translate 翻译模型家族,2B、9B、35B-A3B(preview)三个规模的文本模型权重在 Hugging Face 与 ModelScope 开放,许可证 Apache-2.0([模型卡](https:\u002F\u002Fhuggingface.co\u002FIndexTeam\u002FIndex-Translate-35B-A3B-preview))。\n\n## 35B 总参、3B 激活:翻译专用的 MoE\n\n旗舰 Index-Translate-35B-A3B-preview 基于 Qwen3.5 构建,总参 35B、每 token 激活约 3B——标准稀疏 MoE 配方。覆盖 150 种语言,出厂上下文 262,144 token,官方示例按 32K 档起服务,vLLM 即可自托管。\n\n技术报告披露了三段式流程:先做 167.77B token 的多语种 mid-training(通用\u002F单语\u002F平行语料 constant 阶段 1:1:1,decay 阶段切 1:4:2 的 pivot 组织);再对通用翻译、指令遵循、梗文化翻译三个专家分别 SFT + RL(奖励组合 XCOMET-XXL、语言有效性与 Rubric-as-Reward);最后参数插值合并专家,再用多教师在线蒸馏 MOPD 补弱项([报告](https:\u002F\u002Farxiv.org\u002Fabs\u002F2609.40181))。这套\"专家分工再合并\"的打法,和通用大模型的 post-training 路线已完全同构。\n\n## 把\"约束\"做成硬性规格\n\n真正值得看的是指令遵循。它把翻译约束分两档:硬约束包括术语强制对齐、JSON\u002FCSV\u002Fmarkdown 结构保留、代码块和变量占位符原样保留;软约束负责风格切换(正式、口语、梗味)和领域消歧。对本地化流水线来说,这比\"翻得更像人\"实用得多:术语锁得住,格式不塌,变量不丢。\n\n## 数字怎么看\n\n官方自测表格里,35B-A3B 在报告内所有对比系统中拿到最高 FLORES COMET-22(0.8794)和 instTrans IFscore(0.8336),WMT26 Judge 76.76。同规模对比:腾讯 Hy-MT2-30B-A3B 的 WMT26 Judge 是 66.81,TranslateGemma-12B 是 71.19,Qwen3.5-35B-A3B 底模 71.33——翻译专项训练的增益肉眼可见。低资源语向上 off-target 率仅 2.4%,Hy-MT2-30B-A3B 是 14.5%。\n\n所以要泼冷水:闭源 API 旗舰仍更高,GPT-5.6-Sol 的 WMT26 Judge 是 89.10,DeepSeek-V4.1-Flash 是 83.55。开源翻译模型赢的是自托管、零调用成本与约束可控,不是绝对质量。\n\n## 为什么是 B 站\n\n家族里还有 Index-Echo(语音字幕\u002F语音到语音)、Index-Homura(按目标音节数配音)和 Index-NativeLong(整篇长文档)。组合拳指向很明确:视频字幕、跨语配音、社区内容出海——全是 B 站自己的业务场景。它在梗文化翻译基准 MEME 上拿到 0.7405,明显高于 Hy-MT2-30B-A3B 的 0.5812。UGC 社区的黑话、缩写、玩梗翻译,恰是通用模型最拉胯的角落,而 B 站手里有全世界最密的这类语料。\n\n所以呢:通用模型卷到万亿参数的同时,垂直任务的护城河正变成\"谁有场景数据、谁把约束工程做细\"。B 站这次开源,与其说是慈善,不如说是把业务壁垒顺手变成行业标准件。\n","https:\u002F\u002Fhuggingface.co\u002FIndexTeam\u002FIndex-Translate-35B-A3B-preview","918e1a18-e335-49fb-b11e-026194d7788c",[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},"871127a7-5e83-4e60-b179-08133230ce6a","en","Bilibili open-sources Index-Translate: 35B MoE, 3B active","Bilibili open-sources Index-Translate: a 35B MoE, 3B active, on Qwen3.5, covering 150 languages, Apache-2.0 weights, strong constraint-following.","Machine translation, a task many declared \"solved,\" is getting crowded again. After Tencent Hunyuan and Cohere open-sourced dedicated translation models, Bilibili's Index LLM team delivered its own answer on September 30: the Index-Translate family, with 2B, 9B, and 35B-A3B (preview) text models released under Apache-2.0 on both Hugging Face and ModelScope ([model card](https:\u002F\u002Fhuggingface.co\u002FIndexTeam\u002FIndex-Translate-35B-A3B-preview)).\n\n## A translation-specific MoE: 35B total, 3B active\n\nThe flagship Index-Translate-35B-A3B-preview is built on Qwen3.5: 35B total parameters with roughly 3B activated per token — a standard sparse MoE recipe. It covers text translation across 150 languages, ships with a 262,144-token context window (official examples serve at 32K), and self-hosts on vLLM.\n\nThe technical report describes a three-stage pipeline: first, multilingual mid-training on 167.77B tokens (general, monolingual, and parallel data at 1:1:1 in the constant stage, shifting to a 1:4:2 pivot organization in decay); then specialist SFT and RL for three experts — general translation, instruction following, and meme translation — combining XCOMET-XXL, language-validity, and Rubric-as-Reward signals; finally parameter interpolation merges complementary experts, with multi-teacher on-policy distillation (MOPD) patching remaining weak spots ([report](https:\u002F\u002Farxiv.org\u002Fabs\u002F2609.40181)). The \"divide experts, then merge\" playbook now mirrors mainstream post-training for general LLMs.\n\n## Constraints as a hard spec\n\nThe most interesting part is instruction following. Translation constraints split into two tiers: hard constraints enforce terminology alignment, preserve JSON\u002FCSV\u002Fmarkdown structure, and keep code blocks and variable placeholders intact; soft constraints handle style adaptation (formal, casual, meme-flavored) and domain disambiguation. For localization pipelines this is far more practical than \"sounds more human\": glossaries stay locked, formats don't collapse, variables don't vanish.\n\n## Reading the numbers\n\nOn the vendor's self-reported table, the 35B-A3B preview posts the highest FLORES COMET-22 (0.8794) and instTrans IFscore (0.8336) among all compared systems, with WMT26 Judge at 76.76. Against similarly sized peers: Tencent's Hy-MT2-30B-A3B scores 66.81 on WMT26 Judge, TranslateGemma-12B 71.19, and the Qwen3.5-35B-A3B base only 71.33 — the gains from translation-specific training are visible. On low-resource pairs the off-target rate is just 2.4%, versus 14.5% for Hy-MT2-30B-A3B.\n\nThe cold water: closed API flagships still lead — GPT-5.6-Sol scores 89.10 on WMT26 Judge, DeepSeek-V4.1-Flash 83.55. Open translation models win on self-hosting, zero call cost, and constraint control, not absolute quality.\n\n## Why Bilibili\n\nThe family also includes Index-Echo (speech-to-subtitles and speech-to-speech), Index-Homura (dubbing toward a target syllable count), and Index-NativeLong (full-document translation). The playbook points clearly at video subtitling, cross-language dubbing, and community content going global — all Bilibili's own business scenarios. On the MEME translation benchmark it scores 0.7405, clearly ahead of Hy-MT2-30B-A3B's 0.5812. Slang, abbreviations, and meme-laden UGC text are exactly where general models stumble, and Bilibili holds the densest corpus of it anywhere.\n\nThe takeaway: while general models race toward trillion-parameter scale, the moat for vertical tasks is shifting to \"who has scenario data, and who grinds the constraint engineering fine.\" Bilibili's open-sourcing is less charity than converting its business moat into industry-standard parts.\n","bilibili-index-translate-35b-moe","2026-10-04T13:30:00Z","2026-10-04T13:13:11.530739Z","2026-10-04T13:13:11.530746Z",true,"agent",786,[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},"f3c43720-650b-4054-9349-a1386e06c8fc","Le Chonk 把法国拉回非美\u002F美头部:38 分的 Mistral Large 4","mistral-large-4-le-chonk-intelligence-index-38","2026-10-08T03:30:00+00:00",{"id":65,"title":66,"news_slug":67,"published_at":68},"5ddba2a9-781e-4763-b1e0-1e20c6480391","Mistral Large 4:1万亿参数MoE,月底开源","mistral-large-4-1t-moe","2026-10-06T23:10:00+00:00",{"id":70,"title":71,"news_slug":72,"published_at":73},"0372db20-feaf-42b8-98bd-e42d9c550306","德国Kolibri开源:78B参数只激活3.46B","aleph-alpha-kolibri-1-open-moe","2026-10-03T19:14:02+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"0bc17892-3c47-4081-86a4-3d90afa0c54b","小米 MiMo-V2.6 开源:万亿 MoE 追平 Grok 4.7,Flash 三分之一价格保九成战力","xiaomi-mimo-v2-6-open-weights","2026-09-22T13:02:37+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"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":85,"title":86,"news_slug":87,"published_at":88},"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"]