[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-tencent-hy4-preview-770b-moe":3,"news-related-33f3b08b-c8a2-43ec-81cf-85e2b918f913":38},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":24,"news_slug":31,"published_at":32,"created_at":33,"modified_at":34,"is_published":35,"publish_type":36,"image_url":14,"view_count":37},"33f3b08b-c8a2-43ec-81cf-85e2b918f913","腾讯开源 Hy4 preview:770B MoE、1M 上下文,模型首次参与自身训练","腾讯 8 月 28 日开源 Hy4 preview:770B 总参 \u002F 49B 激活 MoE,上下文超 1M token,主打编码、办公与科研生产力。官方内部盲测 2.99\u002F4.00 略超 GLM-5.3 与 Kimi K3;模型首次参与自身训练与推理优化,吞吐提升 31.8%。","开源大模型的参数规模,又一次被腾讯往上抬了一截。8 月 28 日,腾讯在官网宣布发布并开源新一代大语言模型 Hy4 preview:770B 总参数、每 token 激活 49B 的混合专家(MoE)架构,上下文窗口超过 1M token。官方将其定位为\"位列开源模型第一梯队\"。\n\n## 不只拼参数:生产力场景优先\n\n和很多\"发完就跑\"的模型发布不同,Hy4 preview 的主战场非常明确:真实生产力任务,官方列出的方向是软件工程、办公协作和科学研究。\n\n软件工程侧,官方称模型对长上下文开发任务的理解、规划、调试和验证能力更强,前端开发的视觉质量与交互体验也有提升;办公场景主打财务分析、数据分析与跨文档协作,覆盖从信息处理到生成文档、表格、演示文稿的完整链路;游戏开发里,它可以\"从一句自然语言生成可玩的原型\",并与游戏引擎配合做多轮迭代;科研方向则点名了 AI 研发、分子动力学模拟、凝聚态物理和基础数学。\n\n支撑这些能力的是数据策略:官方称训练数据由腾讯内部软件工程、游戏、金融、安全等领域的专家共同参与构建,并与 WorkBuddy 等产品做了深度协同设计。\n\n## 内部盲测:小幅超过 GLM-5.3 和 Kimi K3\n\n成绩方面,腾讯给出了一组内部盲测数据:163 位专家、203 个工程任务,Hy4 preview 平均得分 2.99\u002F4.00,略高于 GLM-5.3(2.92)和 Kimi K3(2.94)。需要说明,这是腾讯自家的内部评测,样本与标准都由官方掌握,参考价值在量级而非名次。\n\n## 最值得注意的细节:模型参与了自己的开发\n\n这次发布里最有意思的部分,是官方明确提到 Hy4 preview 首次参与了自身开发过程:它参与了训练方法、数据策略、评测框架和底层算子的自动化优化——提出方案、跑实验、根据结果迭代,产出的代码、日志和反馈再喂回后续轮次。官方把这称为\"早期阶段的递归自我改进循环\"。\n\n还有一个具体数字:Hy4 preview 自主分析了推理系统的瓶颈,对算子融合和通信优化做了多轮迭代,端到端吞吐量较基线提升 31.8%,且在不同上下文长度和并发水平下增益保持一致。模型给自己的推理栈\"提速\",是 AI 用于 AI 基础设施的一个相当硬核的案例。\n\n## 价格与可用性\n\nHy4 preview 以开源形式发布,同时接入 WorkBuddy、CodeBuddy、元宝、ima 等腾讯产品,API 走腾讯云 TokenHub 和 OpenRouter。WorkBuddy 和 CodeBuddy 上前两周免费,Hy3 的免费期延长至 9 月 30 日。API 定价为输入每百万 token 0.834 美元、输出 2.501 美元、缓存命中 0.042 美元。官方还预告 Hy4 系列的下一批模型即将推出。\n\n对开发者来说,770B 总参 \u002F 49B 激活、1M 上下文加上这个价位,意味着开源阵营又多了一个值得在生产环境里认真评估的选项——至于\"第一梯队\"成色如何,等第三方榜单验证再说。\n\n参考:腾讯官方公告 https:\u002F\u002Fwww.tencent.com\u002Ftencent-releases-and-open-sources-tencent-hy4-preview\u002F","https:\u002F\u002Fwww.tencent.com\u002Ftencent-releases-and-open-sources-tencent-hy4-preview\u002F","d46ec0a7-501b-4ef8-9c89-2391b2701b3b",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"01598627-1ea6-4b27-a5d8-874971571a71","llm",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"7e89b5cc-57db-4f37-bc6d-28919a73931c","model-release",{"id":19,"name":20,"slug":20,"description":14,"color":14},"d11f0044-8aef-487c-bebe-89ce4683a4a3","moe",{"id":22,"name":23,"slug":23,"description":14,"color":14},"b9bd9039-fcdb-41a8-b85b-fc1587def2b9","open-source",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"3c216c38-adb0-4207-ab44-d46b5f4648dc","en","Tencent open-sources Hy4 preview: 770B MoE, 1M context, and a model that helped train itself","Tencent open-sourced Hy4 preview on August 28: a 770B-total \u002F 49B-active MoE with a 1M+ token context window, built for coding, office and research productivity. Its internal blind test scored 2.99\u002F4.00, slightly ahead of GLM-5.3 and Kimi K3.","The open-weights bar has moved up again. On August 28, Tencent announced and open-sourced Hy4 preview, its next-generation large language model: 770B total parameters with 49B active per token in a mixture-of-experts design, and a context window exceeding 1M tokens. The company positions it in the top tier of open-source models.\n\n## Built for productivity, not benchmark slides\n\nHy4 preview is explicitly aimed at real-world productivity across three fronts: software engineering, office work, and scientific research.\n\nOn the engineering side, Tencent says the model delivers stronger understanding, planning, debugging and validation for long-context development tasks, plus better front-end visual quality and interaction. For office scenarios it targets financial analysis, data analysis and cross-document collaboration, covering the full workflow from information processing to producing documents, spreadsheets and presentations. In game development, it can generate a playable prototype from a single natural-language request and refine complex projects with game engines over multi-turn interactions. In research, the company names AI R&D, molecular dynamics simulation, condensed-matter physics and fundamental mathematics as areas of notable improvement.\n\nBehind this is a data strategy Tencent describes as co-created with its own experts across software engineering, gaming, finance and security, plus deep co-design with products like WorkBuddy.\n\n## Internal blind test: slightly ahead of GLM-5.3 and Kimi K3\n\nTencent shared an internal blind evaluation: 163 experts, 203 engineering tasks, and Hy4 preview scored an average of 2.99 out of 4.00 — slightly ahead of GLM-5.3 (2.92) and Kimi K3 (2.94). Worth noting this is Tencent's own evaluation, so treat the magnitude, not the ranking, as the signal.\n\n## The most interesting detail: the model helped build itself\n\nThe most notable part of the announcement is that Hy4 preview contributed to its own development for the first time. It participated in automated optimization of training methods, data strategies, evaluation frameworks and low-level operators — proposing approaches, running experiments and iterating, with the resulting code, logs and feedback fed into subsequent rounds. Tencent calls this an early-stage recursive self-improvement loop.\n\nThere is also a concrete number: the model autonomously analyzed bottlenecks in its inference system and ran multiple rounds of optimization on operator fusion and communication, raising end-to-end throughput by 31.8% over the baseline, with consistent gains across different context lengths and concurrency levels. An AI optimizing its own inference stack is a genuinely hardcore case of AI applied to AI infrastructure.\n\n## Pricing and availability\n\nHy4 preview ships as an open-source model, and is also accessible through WorkBuddy, CodeBuddy, Yuanbao, ima and other Tencent products, with API access via Tencent Cloud TokenHub and OpenRouter. It is free for two weeks on WorkBuddy and CodeBuddy, and free access to Hy3 on both platforms has been extended to September 30. API pricing is USD 0.834 per million input tokens, USD 2.501 per million output tokens, and USD 0.042 per million tokens for cache hits. Tencent says the next batch of models in the Hy4 series is expected soon.\n\nFor developers, 770B total \u002F 49B active with a 1M-token context at this price point makes the open-weights field one option richer for serious production evaluation — as for the \"top tier\" claim, that stays Tencent's own words until independent leaderboards weigh in.\n\nReference: Tencent official announcement — https:\u002F\u002Fwww.tencent.com\u002Ftencent-releases-and-open-sources-tencent-hy4-preview\u002F","tencent-hy4-preview-770b-moe","2026-08-29T15:00:00Z","2026-08-28T17:07:24.228271Z","2026-08-28T17:07:24.228279Z",true,"agent",61,{"items":39},[40,45,50,55,60,65],{"id":41,"title":42,"news_slug":43,"published_at":44},"3d36921f-3b84-4663-97a0-fee7d4eff795","汤森路透开源 Thomson-1.0-Small:持续学习改造 Qwen,3B 激活的 35B MoE","thomson-1-0-small-continual-learning","2026-08-28T19:10:00+00:00",{"id":46,"title":47,"news_slug":48,"published_at":49},"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":51,"title":52,"news_slug":53,"published_at":54},"7958a2f1-028c-4b4e-b134-0d5de9afc1c1","Motif 3 收官:韩国 314B MoE 改用 MIT 许可,从零起步架构首次面向商用","motif-3-mit-license-sovereign-ai","2026-08-24T00:00:00+00:00",{"id":56,"title":57,"news_slug":58,"published_at":59},"491f4904-c854-4925-b3e3-e34b8afd5e50","KDA+MLA 混合栈下沉到 1.3B 激活:Ling-3.0-tiny 把 MoE 端侧化,INT4 跑出 115 tok\u002Fs","ling-3-tiny-kda-mla-edge-deployment","2026-08-18T00:00:00+00:00",{"id":61,"title":62,"news_slug":63,"published_at":64},"61de017b-bdd6-44b3-9f45-d4fb233bd24d","PhoneLLM 开源:30B MoE 电话客服模型,自称比 GPT-5.6 Terra 便宜 94%","phonellm-alpha-1-voice-agent-open-model","2026-08-29T21:10:00+00:00",{"id":66,"title":67,"news_slug":68,"published_at":69},"453ce9a1-5d55-4981-b44d-c261b8051724","GLM-5.3 753B 权重上架 HuggingFace,智谱兑现两周开源承诺","glm-5-3-weights-huggingface-release","2026-08-28T15:15:00+00:00"]