Models72°

HY4 登陆 Hugging Face:770B 参数、1M 上下文,附带 DFly2 speculator

Why we picked it

770B 加 1M 上下文还带了推测解码,居然连 base 权重和 vLLM 集成都给了。做长文档或自建推理服务的可以看看它的 HY 许可证条款,MaaS 商用有专门限制。

HY4 已在 Hugging Face 上发布,采用 770B 参数规模,搭配一个小型视觉编码器,并配备 DFly2 speculator 用于推测解码。模型支持 1M token 上下文长度,同时放出了 base model 权重。工程层面原生支持 vLLM 和 SGLang 两大主流推理框架的集成。

Body

许可证方面采用 HY License——一份基于 Apache 修改、但针对 MaaS 服务商加入商业条款的协议,发文者推测这意味着该模型可能在韩国和欧盟市场可用。对自建推理服务的团队来说,vLLM/SGLang 的原生支持意味着可以较少改造直接接入现有部署栈,而 base model 的开放也给微调留出了空间。

Original · Xianbao Qian

HY4 is out on @huggingface

I like this chart @ShunyuYao12 . To summarize:

- 770B and a small vision encoder - DFly2 speculator - 1M context length - Base model available too

Some interesting points: - HY license (modified Apache license but with commercial terms for MaaS providers) - meaning that the model is available in Korea and EU now? - Native vLLM and SGLang integration

Link to the model: https://t.co/UOkxU0eq8K

Industry Coordinates & Source Verification

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 HY4, vLLM, SGLang, open-source models.

Source attribution: curated from Xianbao Qian (@Xianbao_QIAN).