Zhipu AI launches third-generation large base model, adapted to domestic chips

·The third-generation large base model ChatGLM3 aims at the visual mode GPT-4V, which improves the ability to understand Chinese pictures and texts, and has enhanced access to search. It can automatically search for relevant information on the Internet based on questions and provide references or references when answering. Article link. The end test models ChatGLM3-1.5B and ChatGLM3-3B support vivo, Xiaomi, Samsung mobile phones and vehicle platforms.

On October 27, at the 2023 China Computer Conference, the Chinese cognitive large model company Beijing Zhipu Huazhang Technology Co., Ltd. (hereinafter referred to as "Zhipu AI") launched the third-generation base large model ChatGLM3, which adopts a multi-stage Enhanced pre-training methods make training more complete, and launched ChatGLM3-1.5B and ChatGLM3-3B that can be deployed on mobile phones, supporting multiple mobile phones and vehicle-mounted platforms including vivo, Xiaomi, and Samsung.

Aiming at the visual mode GPT-4V, ChatGLM3 has implemented iterative upgrades of several new functions, including the multi-modal understanding capability of CogVLM, which has achieved SOTA (the latest in image recognition semantics) on more than 10 international standard image and text evaluation data sets. Best performance, State-of-the-art). The CogVLM model improves the understanding of Chinese graphics and text, can complete complex target detection, and labels it to complete automatic data annotation. Recipes can be given based on photos of ingredients and adapted to the taste of the interlocutor.

Recipes are given based on photos of ingredients.

Zhang Peng, CEO of Zhipu AI, told The Paper (www.thepaper.cn) that multi-modal large models have made a lot of concrete progress in interactive perception of speech, vision, and natural language. In the future, multi-modal large models will The model will move to a more important stage and may integrate more modal data. Multi-modal pre-training will also lead to further improvements in the intelligence or cognitive capabilities of large models.

Can analyze picture content.

ChatGLM3’s code enhancement module Code Interpreter generates and executes code according to user needs, automatically completing complex tasks such as data analysis and file processing. The "code" function currently supports image processing, mathematical calculations, data analysis and other usage scenarios.

Generate code and execute it according to user needs.

Web search enhancement WebGLM access search enhancement can automatically search for relevant information on the Internet based on questions and provide references or article links when answering.

ChatGLM3 integrates AgentTuning technology, activates model agent capabilities, and enables domestic large models to natively support tool invocation, code execution, games, database operations, knowledge graph search and reasoning, operating systems and other scenarios.

Currently, ChatGLM3 has launched end-test models ChatGLM3-1.5B and ChatGLM3-3B that can be deployed on mobile phones. They support a variety of mobile phones and vehicle-mounted platforms including vivo, Xiaomi, and Samsung. They support inference of CPU chips on mobile platforms with a speed of 20 tokens/s.

End test models ChatGLM3-1.5B and ChatGLM3-3B support vivo, Xiaomi, Samsung mobile phones and vehicle platforms.

Zhang Peng said that since the beginning of 2022, the GLM series models launched by Zhipu AI have supported large-scale pre-training and inference on Ascend, Sunway Supercomputing, and Haiguang DCU architectures. At present, Zhipu AI's products have supported more than 10 domestic hardware ecosystems, and joint innovation with domestic chip companies will help the development of domestic native large models and domestic chips.

The above is the detailed content of Zhipu AI launches third-generation large base model, adapted to domestic chips. For more information, please follow other related articles on the PHP Chinese website!