Shangtang releases a new generation of AI big model, with 12 evaluation scores exceeding GPT-4AI | Model | Shangtang
The reporter learned from Shangtang Technology today that the AI language model "Shusheng Puyu" developed by this company in collaboration with multiple research institutions such as Shanghai Artificial Intelligence Laboratory has surpassed GPT-4 in 12 authoritative evaluations, and its comprehensive performance has comprehensively surpassed GPT-3.5 turbo.
In the first half of this year, Shangtang released the "RiRi Ri Xin" large model system and the first domestic base model "Shusheng Puyu" with comprehensive performance surpassing GPT-3.5 turbo. The research and development of this large model has invested approximately 10000 GPUs, and its capabilities have achieved a leapfrog development.
In August of this year, the new model InternLM-123B completed training and increased its parameter count to 123 billion. The language, knowledge, understanding, reasoning, and subject abilities of this new model have significantly improved. On a total of 300000 question sets in 51 well-known evaluation sets worldwide, the overall test score ranks second in the world, surpassing models such as GPT-3.5 turbo and Meta's newly released LLaMA2-70B.
It is reported that InternLM-123 ranks first in 12 major evaluations. Among them, the AGIEval score in the comprehensive evaluation set exam was 57.8, surpassing GPT-4 and ranking first; The evaluation score of CommonSenseQA for knowledge Q&A is 88.5, ranking first; InternLM-123B ranked first in all five reading comprehension evaluations; In addition, ranked first in the five evaluations of reasoning.
Compared with the first generation "Shusheng Puyu" model, InternLM-123B not only generates more accurate and reliable content, but can perform multi-step reasoning and calculation in complex scenes, and also has the ability to self reflect and correct errors. InternLM-123B has also focused on upgrading its code interpreter and plugin calling capabilities. It can use three commonly used tools: Python interpreter, API calling, and search to solve complex tasks and flexibly build AI intelligent agent applications. On this basis, Shangtang's independently developed "discussion" large model will be upgraded to version 3.0 in September.
While launching an attack on the world's leading GPT-4, Shangtang is collaborating with multiple research institutions to promote the construction of an AI big model open source platform. Part of the training data, training code, and base model weights of InternLM-7B have been opened for free to the academic and industrial communities, and support commercial use.
At present, InternLM-7B ranks first in multiple model testing rankings, becoming one of the world's best performing lightweight base models. Shang Tang stated that he looks forward to seeing the AI community improve InternLM and co build more AI applications.
The breakthrough of the big language model has brought new opportunities for the industry. Based on the InternLM model and combined with self-developed inference acceleration algorithms, Shangtang is collaborating with leading mobile chip manufacturers to develop real-time computing capabilities for large language models on the mobile end, which will add a series of new features to the mobile operating system.
"In the first half of this year, big models and generative AI can be said to be the most eye-catching technological breakthroughs in the world." Dr. Xu Li, Executive Chairman and CEO of Shangtang Group, said, "We hope to bring stronger big model capabilities to the industry and help our users make disruptive products in the era of generative AI."
