{ "version": 1, "event_id": "evt_7c6db2b92215084a", "url": "https://xiyu.news/events/evt_7c6db2b92215084a/", "json": "https://xiyu.news/api/events/evt_7c6db2b92215084a.json", "type": "other", "status": "monitoring", "category": "technology", "title": { "zh": "Kolibri 是 Aleph Alpha 推出的面向德语和英语的开源权重 LLM", "en": "Kolibri is an open-weight LLM from Aleph Alpha for German and English" }, "current_state": { "zh": "Aleph Alpha 发布并开源 Kolibri，一个面向德语和英语的 MoE 大语言模型。总参数 781 亿，每 token 激活约 34.6 亿，权重以 Apache 2.0 协议上传 Hugging Face。支持工具调用和可调节推理模式，上下文最高可扩展到约 100 万 tokens（实际训练到 26.2 万）。官方技术报告详细披露数据集与训练方法。", "en": "Aleph Alpha released and open-sourced Kolibri, a German-English MoE LLM. It has 78.1B total parameters and ~3.46B active per token, with weights on Hugging Face under Apache 2.0. It supports tool calling and adjustable reasoning, with context extendable to ~1M tokens (trained to 262K). The official technical report details dataset and training recipe." }, "first_seen_at": "2026-10-03T16:34:57.030858+00:00", "last_updated_at": "2026-10-04T16:56:05.034233+00:00", "last_material_change_at": "2026-10-04T16:56:05.034233+00:00", "confidence": 0.75, "updates_count": 2, "sources_count": 2, "entities": [ "aleph", "aleph-alpha", "alpha", "english", "german", "kolibri", "llm", "open-weight-llm" ], "identifiers": [], "topics": [ "ai-research", "aleph-alpha", "german-language-model", "llm-release", "long-context", "moe-architecture", "open-source-models", "open-weight-llm", "sovereign-ai" ], "updates": [ { "update_id": "upd_d3d72085fc7e8e1a", "event_id": "evt_7c6db2b92215084a", "occurred_at": "2026-10-03T10:43:51Z", "published_at": "2026-10-03T10:43:51Z", "first_seen_at": "2026-10-03T16:34:57.030858Z", "time_precision": "published", "update_type": "initial", "material_change": true, "title_zh": "Kolibri 是 Aleph Alpha 推出的面向德语和英语的开源权重 LLM", "title_en": "Kolibri is an open-weight LLM from Aleph Alpha for German and English", "what_changed_zh": "Aleph Alpha 发布了 Kolibri，这是一个针对德语和英语优化的开源权重（open-weight）大语言模型，并同步公开了一份技术报告，记录了其数据集与训练方法。报告发布于 aleph-alpha.com/downloads/tech-report.pdf。\n\n这是来自欧洲“主权 AI”实验室的一次开源权重发布，评论者特别指出，该技术报告对数据集构建与训练流程的披露异常详尽，开放程度少见。\n\n随附的技术报告被描述为一份“如何从零构建现代 agentic LLM”的逐步教程，其中包括数据集是如何制作的。", "what_changed_en": "Aleph Alpha released Kolibri, an open-weight LLM optimized for German and English, accompanied by a detailed technical report that documents its dataset and training recipe.", "current_state_zh": "Aleph Alpha 发布了 Kolibri，这是一个针对德语和英语优化的开源权重（open-weight）大语言模型，并同步公开了一份技术报告，记录了其数据集与训练方法。报告发布于 aleph-alpha.com/downloads/tech-report.pdf。\n\n这是来自欧洲“主权 AI”实验室的一次开源权重发布，评论者特别指出，该技术报告对数据集构建与训练流程的披露异常详尽，开放程度少见。\n\n随附的技术报告被描述为一份“如何从零构建现代 agentic LLM”的逐步教程，其中包括数据集是如何制作的。", "current_state_en": "Aleph Alpha released Kolibri, an open-weight LLM optimized for German and English, accompanied by a detailed technical report that documents its dataset and training recipe.", "detailed_summary_zh": "Aleph Alpha released Kolibri, an open-weight LLM optimized for German and English, accompanied by a detailed technical report that documents its dataset and training recipe.", "detailed_summary_en": "Aleph Alpha released Kolibri, an open-weight LLM optimized for German and English, accompanied by a detailed technical report that documents its dataset and training recipe.", "background_zh": "Aleph Alpha GmbH 是一家德国 AI 初创公司，开发大语言模型，并强调其结果生成来源的透明度，产品面向欧洲企业与公共机构。所谓开源权重 LLM，是指训练后的参数被公开发布，其许可证通常允许下载、运行、微调乃至商用。“主权 AI”一般指企业或机构自行掌控模型、数据流、质量与运营，而非完全交由外部供应商。", "background_en": "Aleph Alpha GmbH is a German AI startup that develops large language models and says it emphasizes transparency of the sources used to generate its results, with products aimed at European enterprises and public institutions. An open-weight LLM is one whose trained parameters are published under a license that typically permits downloading, running, fine-tuning and commercial use. \"Sovereign AI\" generally refers to retaining control over models, data flows, quality and operations rather than delegating them entirely to an outside provider.", "community_discussion_zh": "Hacker News 讨论帖（306 分、125 条评论）总体上称赞该报告的透明度，有评论者称这是第一次见到如此程度的开放。也有人讨论主权模型应当擅长什么——有人主张其核心作用是以“信任适配器”的方式审计其他模型的输出——还有评论者认为该贴文应提及公司与加拿大企业 Cohere 的合并计划。批评集中在基准对比上：有评论者称，模型选择与约一年前发布的 Qwen3-Next 80B-A3B 对比，却未与 Qwen3.8 Flash 比较，这一点相当刺眼；另有评论者则直接贬低 Aleph Alpha，认为其已未能追上其他实验室。", "community_discussion_en": "The Hacker News thread (306 points, 125 comments) largely praised the report's transparency, with one commenter calling it the first time they had seen this level of openness. Others debated what a sovereign model needs to do — one argued its main job is auditing other models' outputs as a \"trust adapter\" — and a commenter said the post should have mentioned the company's planned merger with Canadian firm Cohere. Criticism focused on benchmarking, with one commenter calling the absence of a comparison to Qwen3.8 Flash striking given the model was instead compared with the roughly year-old Qwen3-Next 80B-A3B, and another dismissing Aleph Alpha as having failed to catch up with rival labs.", "market_impact_zh": "", "market_impact_en": "", "importance_score": 7.5, "references": [ { "url": "https://news.ycombinator.com/item?id=49943034", "title": "Community discussion" }, { "url": "https://en.wikipedia.org/wiki/Aleph_Alpha", "title": "Aleph Alpha - Wikipedia" }, { "url": "https://aleph-alpha.com/en/", "title": "Aleph Alpha" }, { "url": "https://www.zeour.co.uk/glossary/open-weight-llm", "title": "Open-Weight LLM — Llama, Mistral, Qwen, DeepSeek Explained" } ], "confidence": 0.75, "story_ids": [ "hackernews:story:49943034" ], "sources": [ { "url": "https://tej.as/blog/aleph-alpha-kolibri", "label": "tejaskumar__", "source_type": "hackernews", "official": false } ] }, { "update_id": "upd_7d1cfa15645e10f7", "event_id": "evt_7c6db2b92215084a", "occurred_at": "2026-10-04T15:22:02Z", "published_at": "2026-10-04T15:22:02Z", "first_seen_at": "2026-10-04T16:56:05.034233Z", "time_precision": "published", "update_type": "confirmation", "material_change": true, "title_zh": "欧洲开源模型Kolibri来了：3.46B激活参数，支持1M上下文", "title_en": "欧洲开源模型Kolibri来了：3.46B激活参数，支持1M上下文", "what_changed_zh": "新细节：Kolibri 采用 MoE 架构，总参数 781 亿，每 token 激活约 34.6 亿；权重以 Apache 2.0 协议上传 Hugging Face；支持工具调用、可调推理模式；上下文可扩展至约 100 万 tokens（实际训练到 26.2 万）；官方评测 AIME 2025 96.9%、LiveCodeBench v6 85.9%。", "what_changed_en": "New details: Kolibri uses MoE architecture with 78.1B total and 3.46B active parameters per token; weights on Hugging Face under Apache 2.0; supports tool calling and adjustable reasoning; context extendable to ~1M tokens (trained to 262K); official benchmarks AIME 2025 96.9%, LiveCodeBench v6 85.9%.", "current_state_zh": "Aleph Alpha 发布并开源 Kolibri，一个面向德语和英语的 MoE 大语言模型。总参数 781 亿，每 token 激活约 34.6 亿，权重以 Apache 2.0 协议上传 Hugging Face。支持工具调用和可调节推理模式，上下文最高可扩展到约 100 万 tokens（实际训练到 26.2 万）。官方技术报告详细披露数据集与训练方法。", "current_state_en": "Aleph Alpha released and open-sourced Kolibri, a German-English MoE LLM. It has 78.1B total parameters and ~3.46B active per token, with weights on Hugging Face under Apache 2.0. It supports tool calling and adjustable reasoning, with context extendable to ~1M tokens (trained to 262K). The official technical report details dataset and training recipe.", "detailed_summary_zh": "German AI company Aleph Alpha open-sourced Kolibri, an English-German MoE model with 78.1B total and 3.46B active parameters under Apache 2.0, supporting tool calling, adjustable reasoning, and context extendable to roughly 1M tokens (trained to 262K).", "detailed_summary_en": "German AI company Aleph Alpha open-sourced Kolibri, an English-German MoE model with 78.1B total and 3.46B active parameters under Apache 2.0, supporting tool calling, adjustable reasoning, and context extendable to roughly 1M tokens (trained to 262K).", "background_zh": "", "background_en": "", "community_discussion_zh": "", "community_discussion_en": "", "market_impact_zh": "", "market_impact_en": "", "importance_score": 7.0, "references": [], "confidence": 0.95, "story_ids": [ "telegram:theblockbeats:198919" ], "sources": [ { "url": "https://m.theblockbeats.info/flash/370241?from=telegram", "label": "theblockbeats", "source_type": "telegram", "official": false } ] } ] }