[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-tencent-hy3-official-launch":3,"topics-all":36,"news-related-2fb1e49d-6021-4b40-9032-57ccbf95015e":55},{"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},"2fb1e49d-6021-4b40-9032-57ccbf95015e","腾讯混元 Hy3 正式版开源:295B MoE 锁定「Agent 实用性」主战场","7月6日,腾讯混元把 4 月底放出的 Hy3 Preview 打磨成正式版,在 Hugging Face \u002F ModelScope \u002F GitCode \u002F CNB 全渠道以 Apache 2.0 开源,同步上线 Hy3-FP8 量化版。\n\n架构上 Hy3 沿用 295B 总参 \u002F 21B 激活的 MoE 设计,外加 3.8B 的 MTP 层;192 专家中 top-8 激活,80 层主网络 + 1 层 MTP,64 头 GQA(KV 头 8 个),原生 256K 上下文,词表 120,832。属于当下开源旗舰的「标准配置」,但腾讯把重心放到了后训练。\n\n相比 Preview 阶段,正式版最大的变化不在「能不能跑」,而在「能不能上线」——来自 50+ 业务团队的反馈被收敛成三件事:**工具调用稳定性**上,SWE-Bench Verified 上 CodeBuddy \u002F Cline \u002F KiloCode 三种 scaffold 之间的准确率方差被压到 4% 以内,无效调用的死循环显著减少;**抗幻觉**层面,内部评测的幻觉率从 12.5% 降到 5.4%,常识错误率从 25.4% 降到 12.7%;**多轮复杂意图**方面,综合多轮问题率从 17.4% 降到 7.9%,开源长对话基准 MRCR 从 42.9% 跳到 75.1%。\n\n更耐看的是 270 位跨领域专家的盲测:312 组有效对比下,Hy3 拿到 2.67\u002F4,领先 GLM-5.1 的 2.51\u002F4,差距在「前端开发 \u002F CI\u002FCD \u002F 数据存储」三块最明显。换句话说,Hy3 押注的差异化是真实工程链路上的可靠性,而不是刷榜上限。\n\n部署侧官方给了 vLLM 与 SGLang 双 recipe,MTP 推测解码、tool-call\u002Freasoning parser 都是现成的,8 卡 H20-3e 是建议的最小硬件。\n\n**短评**:开源旗舰这半年的竞争,已经悄悄从「谁更长 \u002F 谁奥数更强」转向「谁先把工具调用、长程任务、长上下文这三件事同时做到生产可接受」。Hy3 方向对了,但 FP8 之外能不能把 295B 的推理成本压到「可独立部署」水平,会决定它停留在国产旗舰之列、还是真正成为 Agent 时代的「水电煤」。","https:\u002F\u002Fgithub.com\u002FTencent-Hunyuan\u002FHy3","998df6db-96e6-4b8e-8be1-cfa00a6cd177",[10,14,17,20],{"id":11,"name":12,"slug":12,"description":13,"color":13},"6ad31a14-c0da-42df-81fd-564281f768db","agentic-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},"674d36dd-6f32-47e0-b572-c20722a8ccbb","en","Tencent Hy3 open-sourced: 295B MoE targets agent practicality","On July 6, Tencent Hunyuan polished the Hy3 Preview released in late April into an official version, open-sourcing it under Apache 2.0 across all channels — Hugging Face \u002F ModelScope \u002F GitCode \u002F CNB, with the Hy3-FP8 quantized version also going live simultaneously. Architecturally, Hy3 continues the 295B-total \u002F 21B-activated MoE design, plus a 3.8B MTP layer; top-8 activation among 192 experts, 80 main layers + 1 MTP layer, 64-head GQA (8 KV heads), native 256K context, vocabulary 120,832. It's the \"standard configuration\" of the current open-source flagship tier, but Tencent has put the focus on post-training. Compared to the Preview phase, the biggest change in the official version isn't \"can it run\" but \"can it ship\" — feedback from 50+ business teams has been condensed into three things: on **tool-call stability**, the accuracy variance between CodeBuddy \u002F Cline \u002F KiloCode three scaffolds on SWE-Bench Verified has been compressed to within 4%, with a significant reduction in dead-loop invalid calls; on **hallucination resistance**, internal evaluation's hallucination rate dropped from 12.5% to 5.4%, common-sense error rate from 25.4% to 12.7%; on **multi-turn complex intent**, composite multi-turn question rate dropped from 17.4% to 7.9%, with the open-source long-conversation benchmark MRCR jumping from 42.9% to 75.1%. Even more impressive is the blind test with 270 cross-domain experts: across 312 valid comparisons, Hy3 scored 2.67\u002F4, leading GLM-5.1's 2.51\u002F4, with the gap most pronounced in \"front-end development \u002F CI\u002FCD \u002F data storage\" three blocks. In other words, Hy3's differentiation bet is real-engineering-link reliability, not leaderboard ceilings. On deployment, the official provides dual vLLM and SGLang recipes, MTP speculative decoding, tool-call\u002Freasoning parser all ready to go, with 8× H20-3e as the recommended minimum hardware. **Short take**: the open-source flagship competition over the past half year has quietly shifted from \"who's longer \u002F who's better at Olympiad math\" to \"who can simultaneously get tool-calling, long-horizon tasks, long context into production-acceptable\". Hy3 has the right direction, but whether the FP8 version can compress 295B inference cost to \"independently deployable\" levels will determine whether it stays in the domestic flagship camp, or truly becomes the \"public utility\" of the Agent era.","tencent-hy3-official-launch","2026-07-06T08:05:00Z","2026-07-06T08:07:00.287892Z","2026-08-19T02:08:40.142862Z",true,"agent",227,[37,46],{"slug":38,"tag_slug":38,"title_zh":39,"title_en":40,"intro_zh":41,"intro_en":42,"id":43,"is_active":33,"created_at":44,"modified_at":45},"ai-for-science","AI for Science 2026：从 UniPert 到 GPT-Rosalind 的硬核进化","AI for Science 2026: from UniPert to GPT-Rosalind","生命科学、化学材料、物理世界模型——AI 正在从\"语言工具\"变成\"实验伙伴\"。本专题收录 AI 在三大科学方向的关键节点：UniPert 统一基因与化学扰动空间、GPT-Rosalind 端到端生命科学推理、达摩院 AI 智能体 28 小时找到 4 种超导新材料、Anthropic Claude Science 把工作台做成标准品。","From language tool to lab partner — AI is reshaping life sciences, chemistry\u002Fmaterials, and physical world models. This topic covers the key milestones: UniPert unifying genetic-chemical perturbation spaces, GPT-Rosalind's end-to-end life-sciences reasoning, DAMO's AI agent discovering 4 superconducting materials in 28 hours, and Anthropic's Claude Science workbench going mainstream.","988a4300-5fab-41c4-b5d8-63711a2dc757","2026-09-10T01:34:15.296649Z","2026-09-10T01:34:15.296663Z",{"slug":47,"tag_slug":47,"title_zh":48,"title_en":49,"intro_zh":50,"intro_en":51,"id":52,"is_active":33,"created_at":53,"modified_at":54},"h3-series","MiniMax H3 系列：从开源权重到 35 倍吞吐","MiniMax H3 Series: from open weights to 35x throughput","MiniMax H3 自 2026 年 8 月开源以来节奏密集：官方把生成、参考与编辑收回一个模型；ComfyUI 当天压进 RTX 3060；摩尔线程 3 小时完成国产 GPU 适配；fal 后训练版把吞吐拉到 35 倍；FastH3 蒸馏再砍推理成本。本专题持续追踪 H3 的发布—开源—蒸馏—部署全链路。","Since MiniMax open-sourced H3 in August 2026 the pace has been relentless: one unified omni-modal model, same-day ComfyUI support down to an RTX 3060, a 3-hour Day-0 port to Moore Threads GPUs, fal's post-trained H3 Max at 35x throughput, and FastH3 distillation cutting inference cost further. This topic tracks the full H3 chain — release, open weights, distillation, deployment.","83ef0daa-3c31-4cb3-86ed-e5ee58654d5f","2026-09-08T07:33:19.942193Z","2026-09-08T07:33:19.942209Z",{"items":56},[57,62,67,72,77,82],{"id":58,"title":59,"news_slug":60,"published_at":61},"44aa8908-1cef-48d6-b224-7de12a8d4afd","NeoHorse-1：让 Agent 执行轨迹进入自我改进回路","neohorse-1-agentic-post-training-rsi","2026-09-09T07:19:25+00:00",{"id":63,"title":64,"news_slug":65,"published_at":66},"dcd8b3e1-a3c7-4614-aba4-9002219ea5f6","LibreDB Studio 0.15 发布:本地 LLM 接管数据库交互","libredb-studio-local-llm-agent","2026-09-15T00:00:00+00:00",{"id":68,"title":69,"news_slug":70,"published_at":71},"089f56f3-32ff-4036-89b5-728d5f5a9359","边聊边干活:腾讯混元开源全模态交互 Agent Gander,小脑管对话、大脑管执行","hunyuan-gander-omni-interaction-agent","2026-09-09T21:07:00+00:00",{"id":73,"title":74,"news_slug":75,"published_at":76},"9822a1a7-0014-4bd5-bbe0-492401fe6b96","AllSpark 把搜索 Agent 推到 BrowseComp 88.6:SFT-RL Climbing 与推理时上下文管理","allspark-iris-search-agent-sft-rl-climbing","2026-09-07T07:11:17+00:00",{"id":78,"title":79,"news_slug":80,"published_at":81},"63c30bcd-3ffc-47c5-bd74-c2a9ed8f7c94","DeepSeek Harness 预览版开源:Agent 被拆成可插拔的插件栈,模型只负责想、Harness 负责做事","deepseek-harness-plugin-stack","2026-09-05T06:00:00+00:00",{"id":83,"title":84,"news_slug":85,"published_at":86},"089195fb-7fe5-4ba9-a4bc-8e356fe5e923","BAAI把1000个GitHub仓库蒸馏成5000个技能,科研agent奖牌率31%冲到73%","baai-disco-repo-to-skill-library","2026-09-03T17:07:35+00:00"]