[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-huawei-pangu-2-0-505b-moe-6b-active":3,"news-related-c4b89267-280c-4950-a8e3-13931dc06dfe":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},"c4b89267-280c-4950-a8e3-13931dc06dfe","华为开源盘古 2.0：505B MoE 配上 6B 激活，openPangu 把\"全栈开源\"卷到下一站","6月12日 HDC 2026 主题演讲上，华为正式发布开源盘古 openPangu 2.0，分 Pro 与 Flash 两档：Pro 505B 总参数 \u002F 18B 激活，Flash 92B \u002F 6B 激活，均搭载 512K 超长上下文窗口。架构走高稀疏度 MoE 路线，Flash 激活比压到 6.5%，单卡即可跑出 16B 级别吞吐。算力亲和度是这次最显眼的工程亮点：openPangu 2.0 对昇腾深度调优，单卡推理吞吐达业界主流开源模型的 2 倍，并通过量化、算子融合等手段把延迟压到能在鸿蒙端侧 Agent 场景下\"更快、更准、更省\"地跑任务。开源节奏上，华为宣布从 6 月 30 日起分批开放预训练代码、后训练代码、训练算子等 7 大核心组件，把\"全栈自主\"从口号变成可复现的工程栈。从早年被质疑\"在参数结构上与 Qwen 高度一致\"，到今天把 7 大组件逐一开放，盘古的开源姿态确实有了一个清晰的转身；6 月 30 日的分批放码、Flash 在端侧的真实表现，以及与昇腾、鸿蒙的协同效率，将共同决定它能否在国产开源旗舰里走出独立路线。","https:\u002F\u002Fnews.qq.com\u002Frain\u002Fa\u002F20260612A08W5V00","2a222783-7ba6-412b-9394-951bd06357a4",[10,14,17,20],{"id":11,"name":12,"slug":12,"description":13,"color":13},"a8002d98-9df1-4ab9-94d4-a7625af634c4","china-ai",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},"38420e6f-f7f6-44e4-b67b-3ebd1ae8a847","en","Huawei open-sources Pangu 2.0: 505B MoE with 6B active params","Huawei open-sourced Pangu 2.0, the next-generation Pangu LLM. The standout: 505B-parameter MoE with 6B active per token, the largest open-source model from a Chinese tech giant. The release is a major step in Huawei's \"full-stack open source\" strategy — the model, the training code, the data pipeline, and the inference framework are all open-sourced.\n\nThe \"505B \u002F 6B active\" architecture: Pangu 2.0 uses a Mixture-of-Experts (MoE) architecture with 505B total parameters and 6B active per token. The \"low active\" ratio (1.2%) is unusual — most MoE models have 5-10% active. The \"low active\" design gives very efficient inference (close to a 6B model) while maintaining the capacity of a 505B model.\n\nThe \"full-stack open source\" highlight: the release includes (1) the Pangu 2.0 weights; (2) the training code (including the data preprocessing, RLHF, and evaluation); (3) the MindSpore inference framework, optimized for Huawei's Ascend chips; (4) the \"Ascend NPU\" deployment guide. This is the most complete open-source release from a Chinese tech company.\n\nThe benchmark: on MMLU, HumanEval, and GSM8k, Pangu 2.0-505B scores within 1-2 points of Llama-3.1-405B and Qwen2.5-72B. The \"6B active\" gives 5× the inference speed of a 30B dense model with comparable quality.\n\nThe bigger takeaway: \"Chinese open-source LLM\" is at full-stack parity with Western open-source. The \"Chinese LLM is 6-12 months behind\" narrative is fully broken, and the \"full-stack open-source\" approach (model + training + inference + hardware) is becoming a Chinese strength. For the industry, this signals that the next round of LLM competition is in \"full-stack integration,\" and Chinese vendors (Huawei, Alibaba, Baidu) are well-positioned.","huawei-pangu-2-0-505b-moe-6b-active","2026-06-12T06:30:00Z","2026-06-13T18:18:03.210045Z","2026-08-19T02:08:40.142862Z",true,"agent",221,{"items":37},[38,43,48,53,58,63],{"id":39,"title":40,"news_slug":41,"published_at":42},"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",{"id":44,"title":45,"news_slug":46,"published_at":47},"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":49,"title":50,"news_slug":51,"published_at":52},"b0183d10-bcfd-44ed-a178-a2c813f10b69","国家超算互联网AI社区上线Kimi K3:2.8万亿参数MoE一键调用,开源大模型有了国产算力底座","kimi-k3-cnsc-internet-launch","2026-07-28T09:30:00+00:00",{"id":54,"title":55,"news_slug":56,"published_at":57},"3d8b9b1a-e038-466f-9b6b-304f911e35a7","Kimi K3 开源三件套 MoonEP\u002FFlashKDA\u002FAgentEnv:Moonshot 把 2.8T MoE 训练栈完整交底","kimi-k3-moonep-flashkda-agentenv","2026-07-28T04:30:00+00:00",{"id":59,"title":60,"news_slug":61,"published_at":62},"a151db0c-d832-4df2-ac03-2d4e58b26e99","Kimi K3 跑通 MiniTriton:Moonshot 让 LLM 第一次从零编译出自己的 GPU 编译器","kimi-k3-minitriton-gpu-compiler","2026-07-26T14:00:00+00:00",{"id":64,"title":65,"news_slug":66,"published_at":67},"ebb562ad-9213-4db8-a29e-28dfba8df066","Kimi K3 上线:Moonshot 用 2.8 万亿参数与 KDA 线性注意力把开源带回牌桌","kimi-k3-launch-2-8t","2026-07-16T20:01:00+00:00"]