腾讯 Hunyuan 开源 Hy4 flash 模型,含 FP8 和 Base 版本
腾讯把 Hy4 flash 的权重直接放出来了,224B 激活只有 12B,还带 1M 上下文和多模态。想自建长文档或图文应用的开发者可以下载 FP8 版试试,Base 版也适合拿去继续训练。
腾讯 Hunyuan 将 Hy4 flash 开放权重,一并放出 Hy4 flash-FP8 量化版和 Hy4 flash-Base 基座版,托管在 Hugging Face 和 GitHub。模型总参数 224B、激活参数 12B,采用 MoE 架构,支持 1M 上下文窗口,原生多模态。发布采用 Tencent Hy License,官方预告完整版 Hy4 将在后续推出。FP8 版本方便在显存有限的设备上部署,Base 版本可供研究者做继续训练。12B 的激活参数量意味着推理成本接近小型模型,但总容量保留了大规模 MoE 的能力,适合做日常生产力场景的 API 或自建服务。
这次开源与腾讯此前 Hunyuan 系列的策略一致:主力模型开放权重,同时保留闭源旗舰。对开发者而言,1M 上下文加多模态的组合在同尺寸激活参数的开源模型里少见,可以直接用于长文档处理、代码库理解和图文混合输入的应用。 license 条款需要逐条核对后再商用。
Hy4 flash is now open-weight, including Hy4 flash-FP8 and Hy4 flash-Base. 224B (12B active), 1M context, natively multimodal. Fast, efficient and built for everyday productivity. It's yours now. And stay tuned for Hy4 coming up next! 🤗: huggingface.co/tencent/Hy4-fl… T 🐙 github.com/Tencent-Hunyua… cz � hy.tencent.ai/research/hy4-f… Y0k Released under the Tencent Hy License. See full terms her huggingface.co/tencent/Hy4-fl… OBl📊 62 ⚡
This report is categorized under AITOP's "Models" desk, focusing on China's AI industry and domestic foundation model ecosystem. Within our cross-regional analytical framework, it directly maps to AI developer toolchains, coding agents, and engineering workflows and US-China foundation model benchmarks and commercialization comparison. Key entities and technical topics under continuous tracking include Hunyuan, Hy4 flash, Tencent, open-source models.
Source attribution: curated from Hunyuan (@TXhunyuan). Cross-source corroboration: 11 independent sources have covered this development (including IT之家, Xianbao Qian, pandaily), confirming industry-wide signal strength.