Tencent Hunyuan on November 20thOfficial X accountAnnouncing the weight of the open source flagship model Hy4. according to Hugging Face model card, this time the quantified versions of Hy4 and FP8 Hy4-FP8 and the pre-trained basic model Hy4 base are released. ModelScope and GitCode are simultaneously placed, and a system card containing a security assessment is attached.
The total parameters of the Hy4 language backbone are 770B, each token is enabled 49B, the context length is 1 million tokens, and it can understand text, pictures and videos. Tencent said it open sourced the Hy4 preview On the basis of this, pre-training and post-training have been greatly expanded, visual capabilities have been added, and significant progress has been made in following instructions, long texts, long-process tasks and in-depth research.
The parameter size is the same as preview, with an additional 1.23B visual encoder
Comparing the two versions of the model card, Hy4 is the same 78-layer hybrid expert architecture as preview. The first layer is a dense feed-forward network. The remaining 77 layers each have 256 routing experts plus 1 shared expert. Each token enables 8 routing experts. The difference lies in the addition of a 1.23B parameter visual encoder; the attention module uses sparse attention designed with reference to DeepSeek and Zhipu GLM, and a built-in 3.3B parameter DFly2 draft model for speculative decoding. Tencent said that it can keep the cost of long text inference fixed.
The pre-trained basic model scores listed on the model card, Hy4 base scored 91.6 in MMLU and SimpleQA scored 57.2, which are higher than the 91.1 and 52.1 of DeepSeek-V4-Pro-Base in Tencent's table, but the program question MBPP-plus is 74.0, lower than DeepSeek's 79.1. These are Tencent's own evaluations. Tencent also noted on the model card that the basic model was exposed to a large amount of synthetic instruction data in the middle of training, and occasionally did not answer the questions according to the format during the small sample test, and the scores may be distorted.
Authorization changed from Apache 2.0 to Tencent customized terms
The biggest difference is authorization. Hy4 preview uses Apache 2.0. This time Hy4 uses "Tencent Hy License 1.0" which is rewritten based on Apache 2.0.Licensing termsSection 10 adds commercial restrictions: Companies that use Hy4 to provide APIs and other model-as-a-service (MaaS) services must obtain additional written authorization from Tencent as long as their global revenue exceeds US$50 million within 12 consecutive months, or their products have more than 100 million monthly active users.
The terms also state that those that only limit model functions to end products for specific purposes, or that purely transfer third-party model services, are not considered MaaS; the company's purely internal use, models and output are not provided to third parties, and are not subject to this restriction. The low-priced version launched by Tencent on November 16 Hy4 flash, said at that time that the weights would also be opened in the future.
this article Tencent’s open source flagship model Hy4: 770B parameters, authorization changes to commercial threshold first appeared in Chain News ABMedia。




