Tencent Hunyuan and Yao Shunyu hold Reddit AMA on Hy4, reveal small-model roadmap
On November 21, 2026, Yao Shunyu and the Tencent Hunyuan team held a Reddit AMA about the recently released Hy4 and Hy4 flash models. They revealed that small models are on the roadmap, including plans for models targeting 24GB VRAM, and discussed training details such as MoE sparsity experiments, distillation and reinforcement learning to shorten reasoning chains, and challenges like overthinking and agents finding answers online.
Coverage timeline
量子位量子位
姚顺雨和混元团队上Reddit答网友,Hy小模型已进路线图 衡宇 2026-11-21 14:30:55 来源: 量子位 还透露了不少训练背后的秘诀 听雨 发自 凹非寺 量子位 | 公众号 QbitAI Hy小模型已经在路线图上,面向24GB显存的模型也有计划。 今天凌晨, 姚顺雨和腾讯混元团队 来到Reddit社区,围绕刚发布的Hy4和Hy4 flash做了一场AMA。 AMA是「Ask Me Anything」的缩写,意思是「有什么想问的,尽管问」。网友在评论区提问,混元团队成员直接回帖。 参与名单上只列了Reddit账号,至于哪个马甲是姚顺雨,我们还没对上号…… 这场AMA的主角,是混元本月接连发布的两款模型。 11月4日发布的 Hy4 是旗舰,能理解文本、图像和视频,适合长程软件工程这类需要多步推理、反复调用工具的复杂任务。 11月16日发布的 Hy4 flash ,主打日常办公的速度和成本。它支持图文输入和最长100万Token的上下文,每个Token激活约120亿参数。 Hy4和Hy4 flash都已接入腾讯WorkBuddy。混元团队还把WorkBuddy里的修表、改文档、做PPT等典型任务纳入了内部评测WorkIndex,按这套评测测算,Hy4 flash的平均单任务成本约为旗舰Hy4的1/20。 两款模型的权重都已公开。到了AMA,有网友直接报出自己的电脑配置: 96GB显存、128GB内存,能跑Flash吗? 团队说能跑,主要权重留在GPU上,部分专家层和Engram查表模块可以挪到系统内存。 也有人问llama.cpp,团队说适配正在做。 聊到研发,团队还透露了不少训练背后的秘诀。 Flash的激活参数还能不能减?他们拿做过的MoE稀疏度实验解释取舍。 让Agent自己解题,训练时遇到过什么麻烦?有时它会直接上网找现成答案,跳过解题过程。 还有网友问,后训练里最难改善的行为是什么,团队只回了一个词, Overthinking,过度思考。 旗舰Hy4的推理链有时偏长。到了Flash,团队先从旗舰蒸馏,再用强化学习把推理链压短。 从模型怎么训,到开发者能不能把它跑起来,这场AMA聊了不少平时在模型卡上只有一行、到了真用时却绕不开的问题。 以下是AMA问答节选,按主题编排,中文译文有删节。 Enjoy~ Hy4为什么「想太久」,Flash又是怎么训练

Related stories
Tencent releases Hy4 open-weight model, claims top-tier status, adds vision and computer use
Tencent released Hy4, its latest large language model, on November 4, 2026, claiming it joins the top tier of Chinese open-weight AI for the first time. Hy4 adds vision understanding and computer-use capabilities, with weights to be open-sourced soon and a Flash version planned. Following the release, Tencent announced a tenfold capacity expansion for WorkBuddy due to surging demand, and Hy4 ranked second among Chinese models on Artificial Analysis and first among open-source models on Agent Arena.
5 reports · 4 sources

Tencent open-sources Hy4 flash, 224B-param multimodal model with 12B active parameters
Tencent's Hunyuan team introduced Hy4 flash on November 16, 2026, a native multimodal model with 224B total parameters (including 64B Engram parameters), 12B active per token, and a 1M-token context. The model is now open-sourced on GitHub and HuggingFace, with API pricing at 0.5 yuan per million input tokens and 2 yuan per million output tokens, claiming roughly 1/20th the inference cost of Hy4. It is available across Tencent Cloud TokenHub, WorkBuddy, Yuanbao, WeCom, QQ, ima, and Marvis, with output speeds around 200 tokens per second.
2 reports · 2 sources

Tencent unveils Hunyuan Hy4: 770B-parameter open-weight model with native vision and computer-use
Tencent released Hunyuan Hy4, a 770B-parameter model with 49B active parameters, introducing native vision and computer-use capabilities for the first time. The model is integrated into WorkBuddy and other products, with weights to be open-sourced soon. Hy4 demonstrates high-fidelity visual understanding, such as locating 2-meter roof components in a 2.8-billion-pixel aerial image.
1 reports · 1 sources

Tencent unveils Buddy One cross-app AI collaboration system, Hunyuan Hy4 coming soon
At the 2026 Tencent Global Digital Ecosystem Conference on October 29, Tencent launched Buddy One, a cross-app AI collaboration system for individual users. Buddy One uses WorkBuddy as its workbench, connects products like Tencent Docs, Tencent Meeting, Tencent Drive, and ima, and unifies accounts, points, and cross-product context retrieval via Buddy One ID. Tencent also announced upcoming upgrades to WorkBuddy Enterprise and the imminent release of Hunyuan Hy4 and Hy4 flash.
2 reports · 1 sources

News updates every hour. Subscribe in your RSS reader to get new stories automatically.
Subscribe