[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-ebb562ad-9213-4db8-a29e-28dfba8df066":3},{"id":4,"title":5,"summary":6,"original_url":7,"source_id":8,"tags":9,"published_at":23,"created_at":24,"modified_at":25,"is_published":26,"publish_type":27,"image_url":13,"view_count":28},"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","Moonshot AI 在 WAIC 开幕前夕上线 **Kimi K3**——2.8 万亿参数 MoE,比 DeepSeek V4 Pro(1.6T)大 75%,完整权重计划 7 月 27 日开源。亮点不只是堆参数:K3 首次把 Kimi Delta Attention(混合线性注意力 KDA)与 Attention Residuals 同时落地,叠加 FAST 2025 最佳论文 Mooncake 的 KV-cache 中心化推理栈。\n\n**性能侧 K3 直接撕掉\"开源追闭源\"叙事**:GDPval-AA v2 上 1687 分位列第三,落后 Claude Fable 5 Max(1815)与 GPT-5.6 Sol Max(1747.8);BrowseComp 长程检索 91.2 拿下 SOTA。**最关键演示是 48 小时自主芯片设计**:K3 无人干预跑通 4 mm²、100 MHz、8700 tokens\u002Fs 的自指芯片——把\"长程 Agent\"从榜单拉到可验证任务。\n\n**API $3 \u002F $15 每百万 token、缓存 $0.30**,兼容 OpenAI SDK,1M 上下文自动缓存;三档覆盖 256K–1M 窗口,迁移摩擦砍到接近零。\n\n开源追闭源三年,Kimi K3 用 **2.8 万亿 + 线性注意力 + 长上下文 + 完全开源 + OpenAI 兼容**一次堵上\"上不了生产\"的托词。护城河不再是参数,而是算法、推理栈与生态设计。","https:\u002F\u002Fplatform.kimi.ai\u002Fdocs\u002Fguide\u002Fkimi-k3-quickstart","0ec8f614-42c7-4256-8591-209e1e39eb6b",[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","2026-07-16T20:01:00Z","2026-07-16T20:06:46.980814Z","2026-07-16T20:06:46.980823Z",true,"agent",3]