[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"$f2cw9wbr079ula":3},{"slug":4,"category":5,"publishedAt":6,"titleZh":7,"titleEn":8,"summaryZh":9,"summaryEn":10,"models":11,"hardwares":12,"bodyZh":13,"bodyEn":14,"sourceName":15,"sourceUrl":16,"updatedAt":17},"aleph-alpha-kolibri-1","NEWS","2026-10-04T05:04:00.000Z","Aleph Alpha 开源 Kolibri-1：德英双推理 78B MoE，上下文 1M，原生工具调用","Aleph Alpha open-sources Kolibri-1: 78B German\u002FEnglish reasoning MoE with 1M context and native tool calling","欧洲厂商 Aleph Alpha 发布 Kolibri-1：约 78B 总参数 \u002F 3.46B 激活参数的 MoE 推理模型，德英双语原生训练，上下文最高 1,048,576 tokens，原生支持工具调用，开源协议。","Aleph Alpha released Kolibri-1: a ~78B-total \u002F 3.46B-active MoE reasoning model with native German\u002FEnglish training, up to 1,048,576-token context, and native tool calling, under an open license.",[],[],"### 是什么\n\n10 月 3 日，欧洲 AI 公司 Aleph Alpha 发布开源推理模型 **Kolibri-1**：约 **78B 总参数 \u002F 3.46B 激活参数** 的稀疏 MoE，原生支持德英双语，预训练数据中德语占比显著高于主流英语主导模型。上下文窗口最高 **1,048,576 tokens**，原生支持工具调用。\n\n### 关键设计\n\n- **激活参数仅 3.46B**：推理成本远低于稠密 70B 级模型\n- **德英双语原生**：面向欧洲企业市场，尤其德国制造业、法律与公共部门\n- **1M 上下文**：进入\"百万 token\"梯队\n- **工具调用协议**：与主流 function calling 规范兼容\n\n### 与同类对比\n\n| 维度 | Kolibri-1 | Qwen3.5-35B-A3B | DeepSeek-V4.1-Flash |\n|---|---|---|---|\n| 总 \u002F 激活 | 78B \u002F 3.46B | 35B \u002F 3B | 67B \u002F 3.5B |\n| 上下文 | 1M | 256K | 1M |\n| 双语侧重 | 德\u002F英 | 中\u002F英 | 中\u002F英 |\n| 工具调用 | 原生 | 原生 | 原生 |\n\n### 本地部署角度\n\n激活 3.46B 意味着 Q4_K_M 量化后单卡 24GB 即可跑得动；1M 上下文会显著拉高 KV cache 占用，建议 32GB+ 显存或 CPU 卸载。社区量化版的发布情况本站持续跟踪。","### What it is\n\nOn October 3, European AI company Aleph Alpha released open-source reasoning model **Kolibri-1**: a sparse MoE with ~**78B total \u002F 3.46B active** parameters, native German\u002FEnglish bilingual training (with substantially higher German share than mainstream English-dominated models), up to **1,048,576-token** context, and native tool calling.\n\n### Key design\n\n- **Only 3.46B active**: substantially cheaper inference than dense 70B-class models\n- **Native DE\u002FEN**: targets European enterprise, especially German manufacturing, legal and public sector\n- **1M context**: enters the million-token club\n- **Tool calling protocol**: compatible with mainstream function-calling conventions\n\n### Comparison\n\n| Dimension | Kolibri-1 | Qwen3.5-35B-A3B | DeepSeek-V4.1-Flash |\n|---|---|---|---|\n| Total \u002F Active | 78B \u002F 3.46B | 35B \u002F 3B | 67B \u002F 3.5B |\n| Context | 1M | 256K | 1M |\n| Bilingual focus | DE\u002FEN | ZH\u002FEN | ZH\u002FEN |\n| Tool calling | Native | Native | Native |\n\n### Local deployment angle\n\n3.46B active means Q4_K_M quantization runs on a single 24 GB card; 1M context significantly inflates KV-cache — 32 GB+ VRAM or CPU offload is recommended. Watch community quantized builds in our data section.","AI 日报（ai6s.net）","https:\u002F\u002Fai6s.net\u002F6ac0377c05257b0857147348.html","2026-10-04T01:02:09.710Z"]