A collaborative team from Ant Group, Zhejiang University, University of Liverpool, and East China Normal University presented ScaleOT, a novel cross-domain fine-tuning framework, at AAAI 2025. This framework achieves a 50% improvement in privacy protection and a remarkable 90% reduction in computational cost compared to knowledge distillation, without compromising model performance. ScaleOT offers a highly efficient and lightweight solution for cross-domain fine-tuning of billion-parameter models. The innovative nature of the paper secured it an oral presentation at AAAI.

The 360Zhiniao team, in collaboration with Peking University, officially released the medium-sized inference model Tiny-R1-32B-Preview. This model, with only 5% of the parameters, successfully approaches the full performance of Deepseek-R1-671B, demonstrating the enormous potential of small models in efficient inference. The model's performance is particularly outstanding in several key areas. In mathematics, Tiny-R1-32B-Preview...

Recently, Meta has announced a groundbreaking research achievement by developing a new type of memory layer technology that can significantly enhance the factual accuracy of large language models (LLMs) and achieve unprecedented expansion in parameter scale. This technology not only challenges traditional neural network expansion methods but also provides new directions for future AI architecture design. The core of this research lies in utilizing a trainable key-value lookup mechanism to add extra parameters to the model without increasing computational load (FLOPs). This method's core thinking is...

Recently, State Grid Corporation announced the launch of China's first trillion-parameter artificial intelligence large model in the power industry - the 'Bright Power Large Model'. They have signed a strategic cooperation framework agreement with Baidu and Alibaba. The officials stated that they will work together with the signing parties to build the Bright Power Large Model and promote the integrated development of energy and power technological innovation and industrial innovation.

With the rapid development of artificial intelligence, the integration of visual and language capabilities has led to groundbreaking advancements in visual language models (VLMs). These models aim to simultaneously process and understand both visual and textual data, being widely applied in scenarios such as image captioning, visual question answering, optical character recognition, and multimodal content analysis. VLMs play a significant role in developing autonomous systems, enhancing human-computer interaction, and creating efficient document processing tools, successfully bridging the gap between these two data modalities. However, there are challenges in handling high-resolution visual data and diverse textual input.

Recently, Jiyue Xingchen announced the launch of a new member in the Step series model matrix—the Step-1o end-to-end speech model, claimed to be the first trillion-parameter end-to-end speech model in China. The release of this new model marks a significant advancement in the field of speech technology. By utilizing an end-to-end speech solution, it integrates speech understanding and generation, thereby enhancing the IQ and EQ limits of the model and providing a supernatural, high-quality communication experience.

In the arena of artificial intelligence, China is steadily advancing with a posture of independent innovation. On November 26, the world's first multimodal large model with 100 billion parameters, 'Zidong Taichu' 3.0, was officially unveiled in Wuhan, adding a significant chapter to domestic AI technology. This is not just a technological iteration, but an important milestone in China's AI ecosystem. Jointly developed by the Institute of Automation, Chinese Academy of Sciences, Huawei, and the Wuhan East Lake High-Tech Zone, 'Zidong Taichu' has successfully built the world's first pre-trained model covering image, text, and audio modalities.

During the 2024 World Internet Conference held today in Wuzhen, Honor CEO Zhao Ming shared his vision for the future development of mobile phones. He stated that by 2027, phones are expected to be able to incorporate large AI models with 100 billion parameters. This advancement will enable phones to learn, understand, and predict user behaviors and intentions, thereby providing users with more personalized arrangements for their lives and work. Image source note: Image generated by AI, image authorized by service provider Midjourney. Currently, several phones on the market are equipped with models of similar scale.

Recently, a startup in Seattle named Moondream launched a compact visual language model called moondream2. Despite its small size, the model has shown outstanding performance in various benchmark tests and has garnered significant attention. As an open-source model, moondream2 is expected to enable local image recognition capabilities on smartphones. Moondream2 was officially released in March and can process both text and image inputs, possessing functions such as answering questions, text extraction (OCR), object counting, and item classification.

Recently, research teams from several Chinese scientific institutions have launched the super large-scale multimodal dataset named Infinity-MM, and based on this dataset, they have trained an outstanding AI model called Aquila-VL-2B. This breakthrough injects new momentum into the development of multimodal AI. The Infinity-MM dataset is astonishing in size, containing four major types of data: 10 million image descriptions, 24.4 million general visual instruction data, 6 million selected high-quality instruction data, and 3 million generated by G