Tencent Hybrid Large Model Multimodal Transition and Future Plans

TapTechNews July 5th news, according to Interface News, at today's 2024 World Artificial Intelligence Conference (WAIC2024), Jiang Jie, the vice president of Tencent Group and the person in charge of the Tencent Hybrid Large Model, talked about issues related to multimodal of large models.

Jiang Jie stated that the large model industry is transitioning from the initial single-modal to multimodal. He believes that for the Tencent Hybrid Large Model, multimodal is a'must-answer question', and currently the Hybrid Large Model is actively deploying from multimodal to full-modal technology, and users will soon be able to experience it in Tencent Yuanbao App, Tencent's internal businesses and scenarios, and at the same time will be opened to external applications through Tencent Cloud.

And on May 17th, Jiang Jie once revealed at the Tencent Generative AI Industry Application Summit that in the third quarter of this year, the'most core' text-generation model of Hybrid will be open-sourced. Tencent has begun to train three different sizes (S, M, L) of text-generation models for open source use.

The S version is mainly deployed in mobile phones and has a parameter count of 3B;

The M version is mainly deployed in PCs and has a parameter count of 5B;

The L version is mainly deployed in cloud/data centers and is a 30B parameter count MOE architecture model.

According to TapTechNews' previous report, the Tencent Hybrid text-to-image large model (Hybrid DiT) announced an open-source small video memory version yesterday, only requiring 6GB of video memory to run, which is relatively friendly to developers who use personal computers for local deployment.