The Shanghai Artificial Intelligence Laboratory has announced a significant version upgrade for its Shusheng large model, introducing Shusheng ・ Puyou 3.0 (InternLM3). According to the laboratory, the new version has significantly improved data usage efficiency through a refined data framework, resulting in enhanced cognitive density.
The upgraded InternLM3-8B-Instruct model was trained using only 4T of data, and the officials stated that its overall performance exceeds that of other open-source models of comparable size, with training costs reduced by over 75%. Notably, this version is the first to achieve a fusion of regular conversation and deep thinking abilities in a general model, enabling it to better handle diverse real-world usage scenarios.
In terms of model evaluation, the research team employed a unified and reproducible method based on the Sinan OpenCompass open evaluation framework. The evaluation covered more than a dozen authoritative assessment sets, including CMMLU and GPQA, encompassing various dimensions such as reasoning, mathematics, programming, instruction following, long text generation, dialogue, and overall performance. The evaluation results indicate that Shusheng ・ Puyou 3.0 scores ahead in most assessment sets, with overall performance very close to GPT-4o-mini.
The Shanghai AI Laboratory also mentioned that this new version of the model is the first general dialogue model in the open-source community to support browser usage, capable of supporting over 20 steps of web navigation, thus enabling deep information mining.
Experience page: https://internlm-chat.intern-ai.org.cn.
Key Points:
🌟 The Shusheng ・ Puyou 3.0 model is trained on 4T of data, with overall performance surpassing other open-source models of similar scale, achieving over 75% cost savings in training.
📊 The model scores lead in multiple authoritative assessment sets, significantly enhancing the fusion of cognitive and conversational abilities.
🌐 The new model supports browser usage, enabling deep information mining and becoming a highlight in the open-source community.