[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-tencent-hunyuan-hy3-preview-295b-moe":3,"news-related-010979e2-4e0a-4dcc-8a91-c6a99bfebc04":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},"010979e2-4e0a-4dcc-8a91-c6a99bfebc04","腾讯混元 Hy3 Preview 开源：295B MoE 剑指 Agent 实用性","4月23日，腾讯发布并开源混元 Hy3 Preview——这是其2026年2月完成预训练与强化学习基础设施重建后推出的首个模型。295B 总参数、21B 激活参数、256K 上下文，核心目标：不再执着 benchmark 刷分，转而剑指 Agent 能力与真实场景落地。Hy3 采用快慢思考融合的混合专家（MoE）架构。这种设计并非腾讯首创——Mixtral、DeepSeek V3 都验证过 MoE 推理效率的优势——但这次的重点在于融合而非单纯堆参数：慢思考处理复杂推理，快思考响应日常任务，21B 激活参数在保证能力的同时控制了推理成本。256K 上下文支持则为复杂 Agent 任务提供了基础。腾讯首席AI科学家姚顺雨强调，他们希望通过自建评测体系规避传统榜单刷分陷阱。这个表态背后是行业共识的变化：头部厂商开始强调真实场景表现而非榜单第一，是行业走向成熟的信号。API 定价低至 1.2 元\u002F百万 Tokens，性价比成为直接卖点。目前 Hy3 已在腾讯云、元宝、QQ、腾讯文档等产品上线，同时支持 OpenClaw、KiloCode 等开源 Agent 框架。与其在通用 benchmark 上硬碰硬，不如在代码生成、Agent 等场景建立开发者生态护城河——这是腾讯的差异化策略。混元重建计划的第一步，能否兑现，社区反馈会是试金石。","https:\u002F\u002Fcloud.tencent.com\u002Fdeveloper\u002Farticle\u002F2660040","d46ec0a7-501b-4ef8-9c89-2391b2701b3b",[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},"b1853a5a-d940-42b7-94f9-0488ee3f2cf7","new-model",{"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},"14cf54d8-f4d7-4510-b091-7219ff3220c7","en","Tencent Hy3 Preview open-sourced: 295B MoE for real agents","On April 23, Tencent released and open-sourced Hunyuan Hy3 Preview — its first model after completing the pretraining and reinforcement learning infrastructure rebuild in February 2026. 295B total parameters, 21B activated parameters, 256K context, the core goal: no longer obsessing over benchmark scores, but targeting Agent capability and real-scenario deployment. Hy3 uses a fast-slow thinking fused Mixture-of-Experts (MoE) architecture. This design isn't Tencent's invention — Mixtral and DeepSeek V3 have both validated MoE inference efficiency's advantages — but this time the focus is on fusion rather than simply stacking parameters: slow thinking handles complex reasoning, fast thinking responds to daily tasks, with 21B activated parameters controlling inference cost while preserving capability. 256K context support provides the foundation for complex Agent tasks. Tencent's chief AI scientist Yao Shunyu emphasized that they hope to avoid traditional leaderboard gaming by building their own evaluation system. This statement reflects a change in industry consensus: leading vendors beginning to emphasize real-scenario performance over leaderboard first place is a signal of the industry's maturation. API pricing is as low as 1.2 yuan per million tokens, with cost-performance as a direct selling point. Hy3 has already been deployed on Tencent Cloud, Yuanbao, QQ, Tencent Docs, and supports open-source Agent frameworks like OpenClaw and KiloCode. Rather than colliding head-on on general benchmarks, building developer ecosystem moats in code generation and Agent scenarios — that's Tencent's differentiation strategy. Whether the first step of Hunyuan's rebuild plan can be delivered, community feedback will be the litmus test.","tencent-hunyuan-hy3-preview-295b-moe","2026-05-08T13:00:00Z","2026-05-08T13:09:43.360530Z","2026-08-19T02:08:40.142862Z",true,"agent",128,{"items":37},[38,43,48,53,58,63],{"id":39,"title":40,"news_slug":41,"published_at":42},"88d40bcd-f92f-481e-b644-5da3dd9813a4","四家中国实验室十二天内密集发布开源代码模型：前沿能力与低成本并行","4-china-labs-12-days-coding-open","2026-05-21T02:30:00+00:00",{"id":44,"title":45,"news_slug":46,"published_at":47},"6836f7c8-c430-4722-afaa-35d95f40e100","开源LLM的崛起：从追赶引领到标准制定","open-source-llm-rising-china-qwen-glm-deepseek","2026-04-25T02:05:00+00:00",{"id":49,"title":50,"news_slug":51,"published_at":52},"fc0cd2d4-cc5f-4f97-b91f-08719e41e8ec","Qwen3.6-27B：27B密集模型超越397B MoE，单卡部署的编程新选择","qwen-3-6-27b-dense-beats-397b-moe-coding-77pct","2026-04-24T03:30:00+00:00",{"id":54,"title":55,"news_slug":56,"published_at":57},"d4d40e4b-04e9-45cf-a04c-792aca45b152","Kimi K2.6开源发布：万亿参数MoE模型的长时编程与Agent Swarm突破","kimi-k2-6-trillion-moe-1t-32b-active-256k","2026-04-24T03:00:00+00:00",{"id":59,"title":60,"news_slug":61,"published_at":62},"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":64,"title":65,"news_slug":66,"published_at":67},"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"]