Tencent Releases and Open-Sources Tencent Hy4 Preview

Tencent Releases and Open-Sources Tencent Hy4 Preview

摘要中文导览（来自条目评分时的双语摘要，基于原文提炼）：

• 腾讯发布并开源混元新一代模型 Hy4 preview：总参数 770B、激活 49B（MoE）、上下文超 1M token，定位编码、办公与科研等真实生产力任务；经 WorkBuddy/CodeBuddy、元宝、ima 落地，可走腾讯云 TokenHub 与 OpenRouter API，上线两周限时免费，Hy3 免费期延至 9 月 30 日。官方口径：163 名内部专家、203 个工程任务的盲测中 Hy4 preview 均分 2.99/4，GLM 5.3 为 2.92、Kimi K3 为 2.94——差距在噪声量级且为自家组织内部评测。更有信息量的是两条自研过程披露：这一代模型首次参与自己训练方法、数据策略、评估体系与底层算子的自动优化（模型提方案、跑实验、拿结果继续迭代）；以及它自主分析推理系统瓶颈后围绕算子融合与通信优化多轮调整，端到端吞吐较基线提升 31.8% 且在不同上下文长度与并发下稳定。“递归自我改进闭环”是官方自我定性，31.8% 未给出基线细节。

文章信息

• 作者: Tencent
• 发布: 2026-08-28
• 原文链接: https://www.tencent.com/tencent-releases-and-open-sources-tencent-hy4-preview/
• 标签: open-weights tencent moe long-context self-improvement

原文摘录（开头）

_Ranked among the top tier of open-source models, Hy4 preview is built for real-world productivity tasks, delivering outstanding performance across coding, office work, and scientific research_

Tencent has released and open-sourced Tencent Hy4 preview, a next-generation large language model with 770B total parameters and 49B active parameters, and a context window exceeding 1M tokens. It demonstrates outstanding capabilities on real-world productivity tasks spanning coding, office work, and scientific research.

Hy4 preview is now available as an open-source model and can also be accessed globally through WorkBuddy and CodeBuddy, as well as Yuanbao, ima and other Tencent products. Users can try the model directly through these applications, or connect to it via API through Tencent Cloud TokenHub and OpenRouter.

Upon launch, Hy4 preview will be available for free on WorkBuddy and CodeBuddy for two weeks. Free access to Hy3 on both platforms has also been extended until September 30.

Hy4 preview was expanded significantly in model size, context length, and data volume, and the advances in both pre-training and post-training have led to a major leap in overall intelligence, placing the model among the top tier of open-source models.

Summary (EN)

Tencent released and open-sourced Hy4 preview: 770B total / 49B active parameters (MoE), 1M+ token context, positioned for coding, office, and research productivity tasks, available via WorkBuddy/CodeBuddy, Yuanbao, ima, Tencent Cloud TokenHub, and OpenRouter with a two-week free launch window. Official blind evaluation by 163 internal experts on 203 engineering tasks scored 2.99/4 versus GLM 5.3 at 2.92 and Kimi K3 at 2.94 — noise-level margins in a self-organized internal evaluation. More substantive are two disclosures: the model participated for the first time in automating its own training methods, data strategy, evaluation system, and operator optimization, and it autonomously analyzed inference-system bottlenecks, delivering a claimed 31.8% end-to-end throughput gain via operator fusion and communication optimizations (baseline details not provided).

Obsidian 证据摘录

入选自 Obsidian《ClawFeed 24小时高价值一览 · 2026-08-31》第5篇：官方一手发布稿，含自研过程披露与 31.8% 吞吐数字。