The Shanghai Artificial Intelligence Laboratory recently announced that its self-developed scholar model has received a major upgrade and launched a new version of Scholar Puyu 3.0 (InternLM3). This version has made significant breakthroughs in data usage efficiency and model performance, achieving higher performance at a lower cost, and for the first time integrating regular dialogue and in-depth thinking capabilities, significantly improving the model's performance in real application scenarios . This upgrade not only achieves a technological breakthrough, but also reflects China’s determination and strength to continue to innovate in the field of artificial intelligence.
Shanghai Artificial Intelligence Laboratory announced that its scholar model has received an important version upgrade and launched Scholar Puyu 3.0 (InternLM3). According to the laboratory, the new version has significantly improved data usage efficiency through a refined data framework, thus achieving an increase in thinking density.
The upgraded InternLM3-8B-Instruct model only uses 4T of data for training. Officials say its comprehensive performance exceeds that of open source models of the same size, and training costs are saved by more than 75%. It is worth noting that this version for the first time achieves the integration of regular dialogue and in-depth thinking capabilities in a general model, which can better cope with diverse real-life usage scenarios.
In terms of model evaluation, the research team adopted a unified and reproducible method for evaluation based on the Sinan OpenCompass open source evaluation framework. The evaluation content involves more than ten authoritative evaluation sets such as CMMLU and GPQA, covering multiple dimensions such as reasoning, mathematics, programming, instruction following, long text generation, dialogue and comprehensive performance. The evaluation results show that Shusheng Puyu 3.0 leads the score in most evaluation sets, and its overall performance is very close to GPT-4o-mini.
The Shanghai AI Laboratory also stated that this new version of the model has become the first universal dialogue model in the open source community to support browser use, and can support web page jumps of more than 20 steps, thereby enabling the mining of in-depth information.
Experience page: https://internlm-chat.intern-ai.org.cn.
Highlight:
The Shusheng Puyu 3.0 model is trained with 4T data, and its comprehensive performance exceeds that of open source models of the same scale, saving more than 75% of training costs.
The model scores leading in multiple authoritative evaluation sets, and the integration of thinking and dialogue capabilities has been greatly improved.
The new model supports browser use and can conduct in-depth information mining, becoming one of the highlights of the open source community.
All in all, the upgrade of Shusheng Puyu 3.0 demonstrates China’s significant progress in the field of large-scale language models. Its efficient training methods and powerful performance are expected to promote the application of artificial intelligence technology in more fields, and it is worth looking forward to its future development.