[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-cohere-north-mini-code-30b-3b-h100":3,"news-related-747713d0-690f-45c7-afb0-7d6e16cb2a33":36},{"id":4,"title":5,"summary":6,"content":6,"original_url":7,"source_id":8,"tags":9,"translations":23,"news_slug":29,"published_at":30,"created_at":31,"modified_at":32,"is_published":33,"publish_type":34,"image_url":13,"view_count":35},"747713d0-690f-45c7-afb0-7d6e16cb2a33","Cohere North Mini Code 开源：30B MoE、3B 激活，单卡 H100 跑起 Agentic Coding","Cohere 把\"主权 AI\"押在了开发者身上——North Mini Code 1.0 是其第一款专为 Coding Agent 设计的开源模型，30B 总参数 \u002F 3B 激活 MoE、Apache 2.0 授权、最低 1× H100 FP8 即可本地跑。\n\n## 模型规格\n\n- 30B 总参 \u002F 3B 激活的 Mixture-of-Experts 结构\n- 256K 总上下文窗口，64K 最大生成长度\n- 训练目标：代码生成、Agentic 软件工程、Terminal 任务\n- 显式兼容 OpenCode 等主流 Agent Harness\n\n## 性能与吞吐\n\n- Artificial Analysis Coding Index 33.4 分，在同尺寸开源模型中靠前\n- 内部测试下，输出吞吐量比 Devstral Small 2 高 2.8×\n- Inter-token latency 优于 Devstral Small 2 约 30%\n- SWE-Bench Verified \u002F Pro 与 Terminal Bench v2 \u002F Hard 均有公开对比\n\n## 为什么这件事值得关注\n\n30B 总参 + 3B 激活 + 256K 上下文 + Agentic 训练，正在成为 2026 年下半年 Coding Agent 模型的\"标准配方\"：大参数容量兜住长代码库语义，小激活参数守住推理成本，长上下文让 Agent 能扫完整仓库，专项训练则补足多步工具调用的一致性。North Mini Code 把这条路线上的每一项都做到了当前开源的最优解之一。\n\n更关键的是单卡 H100 即可私有化部署。对代码安全敏感、又不想被任何闭源 Coding Agent 锁定工作流的厂商来说，这是少有的\"既能本地跑、又有 Apache 2.0 可改\"的选择。Hugging Face 上的权重显示，自 6 月 5 日开放早期访问以来已经积累了上千次下载，社区反应也偏正面。\n\n但 MoE 的路由效率和小激活参数的稳定性，要在真实 IDE 与 CI 场景里扛过几周工作流才见真章。Anthropic、OpenAI、Google 都在把 Coding Agent 当作下一阶段 LLM 落地的核心场景，Cohere 这次以\"开源 + 小激活 + 主权\"切入，是对开源生态的一次精准补位——也是它重新拿回开发者注意力的关键一步。","https:\u002F\u002Fcohere.com\u002Fblog\u002Fnorth-mini-code","df9f8204-8e8d-4fce-8526-3c6fe8e6ae56",[10,14,17,20],{"id":11,"name":12,"slug":12,"description":13,"color":13},"e82b2d09-81b2-43d1-977e-e018443b3c14","coding-agent",null,{"id":15,"name":16,"slug":16,"description":13,"color":13},"01598627-1ea6-4b27-a5d8-874971571a71","llm",{"id":18,"name":19,"slug":19,"description":13,"color":13},"7e89b5cc-57db-4f37-bc6d-28919a73931c","model-release",{"id":21,"name":22,"slug":22,"description":13,"color":13},"b9bd9039-fcdb-41a8-b85b-fc1587def2b9","open-source",[24],{"id":25,"lang":26,"title":27,"summary":28,"content":13},"c6499c53-2048-4b35-b408-74c9d347a4dd","en","Cohere North Mini Code: 30B MoE agentic coding on one H100","Cohere is betting \"sovereign AI\" on developers — North Mini Code 1.0 is its first open-source model designed specifically for Coding Agents, with 30B total \u002F 3B active parameters in a MoE layout, Apache 2.0 license, runnable locally on as little as 1× H100 FP8.\n\n## Model specs\n\n- 30B total \u002F 3B active Mixture-of-Experts structure\n- 256K total context window, 64K max generation length\n- Training objectives: code generation, Agentic software engineering, Terminal tasks\n- Explicitly compatible with mainstream Agent Harnesses like OpenCode\n\n## Performance and throughput\n\n- Artificial Analysis Coding Index 33.4, ahead of peers at the same size among open-source models\n- In internal testing, output throughput is 2.8× higher than Devstral Small 2\n- Inter-token latency about 30% better than Devstral Small 2\n- Public comparisons available on SWE-Bench Verified \u002F Pro and Terminal Bench v2 \u002F Hard\n\n## Why this matters\n\n30B total + 3B active + 256K context + Agentic training is becoming the \"standard recipe\" for Coding Agent models in the second half of 2026: large total parameters carry long-codebase semantics, small active parameters guard inference cost, long context lets the Agent scan the full repo, and specialized training patches up multi-step tool-use consistency. North Mini Code nails every one of these to the current open-source best.\n\nWhat's even more important is the single-H100 deployability. For vendors that care about code security and don't want to be locked into any closed-source Coding Agent workflow, this is one of the few \"runs locally AND is modifiable under Apache 2.0\" options. The Hugging Face weights show that since early-access opened on June 5, the model has accumulated over a thousand downloads, with community reaction skewing positive.\n\nWhether MoE routing efficiency and small-active-parameter stability hold up in real IDE and CI workflows for a few weeks of real work is the real test. Anthropic, OpenAI, and Google are all treating Coding Agents as the core scenario for the next stage of LLM deployment, and Cohere's entry into \"open-source + small-active + sovereign\" is a precise gap-fill for the open-source ecosystem — and a key step in regaining developer mindshare.","cohere-north-mini-code-30b-3b-h100","2026-06-11T12:00:00Z","2026-06-11T12:15:25.558828Z","2026-08-19T02:08:40.142862Z",true,"agent",143,{"items":37},[38,43,48,53,58,63],{"id":39,"title":40,"news_slug":41,"published_at":42},"89a79f9a-bfd2-4ebe-8f03-92fa74a3a34f","Ornith-1.5 开源：模型自己出题、自己搭考场，397B 到 9B 三档齐发","ornith-1-5-self-improvement-open-models","2026-08-20T13:30:00+00:00",{"id":44,"title":45,"news_slug":46,"published_at":47},"1e7d0673-aecc-42b5-8560-92a2b4d4daf6","快手 KAT-Coder-V2.5 把 Agentic Coding 训练改写成基础设施工程","kuaishou-kat-coder-v2-5","2026-07-27T06:00:00+00:00",{"id":49,"title":50,"news_slug":51,"published_at":52},"6ceaf229-f1a2-4231-b531-797a99faa194","Ornith-1.0：模型自写 RL harness，SWE-Bench 比肩 Opus 4.7","ornith-1-0-397b-moe-swe-bench-opus-4-7","2026-06-26T18:01:01+00:00",{"id":54,"title":55,"news_slug":56,"published_at":57},"d63524e2-85bc-484c-b4e7-7fac32c3ac08","GLM-5.2 即将全量上线 Coding Plan：智谱把\"编程开源\"卷成新一轮标配","glm-5-2-coding-plan-zhipu-open-source","2026-06-13T07:30:00+00:00",{"id":59,"title":60,"news_slug":61,"published_at":62},"4d436945-18e9-4d69-a4c8-c1e3e975ab33","MiniMax M3发布：稀疏注意力打通百万token上下文，开源模型编程能力逼近闭源前沿","MiniMax-m3-sparse-attn-million-token-msa","2026-06-04T01:00:00+00:00",{"id":64,"title":65,"news_slug":66,"published_at":67},"f6e4aab0-7693-4c2c-bb66-c1641fc2cc3e","Ox Alpha 谜底揭晓:智谱 GLM-5.3-Flash,MIT 开源 320B MoE","ox-alpha-glm-5-3-flash-reveal","2026-08-27T13:30:00+00:00"]