AI2 Releases Open Source Dataset for Large Language Model Dolma Containing 3 Trillion Tokens

The Allen Institute for Artificial Intelligence in the United States recently released an open-source dataset called Dolma, which contains 3 trillion tokens. This dataset will serve as the foundation for the Open Language Model (OLMo) being developed by AI2, with a planned release in early 2024. The data in Dolma is sourced from a wide range of materials, including web content, academic publications, code, and books. This dataset is currently the largest of its kind available publicly.

Qwen 3.6 Officially Released: 1 Million Long Context, Competing with Claude Code

Alibaba released the new generation large language model Qwen3.6-Plus, which is hailed as the strongest domestic programming model at present. Compared to the 3.5 version, its performance has been significantly improved, ranking first among domestic models in multiple programming evaluations, and its overall capabilities are close to the international benchmark Claude series. The model demonstrates a high level of autonomy in front-end development, complex repository tasks, and other areas.

ChatGPT-5.2 First Independently Proves a Mathematical Conjecture, Opening the Era of Contextual Proof

Research from the Free University of Brussels found that commercial large models are now capable of independently generating original mathematical proofs. ChatGPT-5.2 successfully solved a mathematical conjecture proposed in 2024, marking a milestone where the capabilities of large language models have surpassed code assistance and text creation, entering the field of mathematics requiring strong logical reasoning.

Unsloth Studio Launches: The First Local Visual Large Model Fine-tuning Platform, Reducing VRAM Usage by 70%

Unsloth AI releases an open-source no-code visual tool called Unsloth Studio, aiming to simplify the fine-tuning process of large language models and lower the development threshold. The tool achieves double the training speed and saves 70% of VRAM usage through a customized backpropagation kernel, without requiring complex environment configuration or high hardware costs.

Carnegie Mellon University Develops New AI System: Real-Time Repair of 3D Printing Defects Like a Conductor

A team from Carnegie Mellon University has developed a real-time error correction system for 3D printing based on a large language model. The system mimics an orchestra, with an 'conductor' agent coordinating four specialized agents to automatically detect and correct errors caused by small parameter fluctuations during the printing process, solving the problem of traditional open-loop systems that are prone to failure.

Neusoft Group Collaborates with Cerence AI to Create a New Experience for the Future Intelligent Cockpit

Neusoft Group has reached a strategic cooperation with Cerence AI to jointly develop a new generation of intelligent cockpit platform, focusing on intelligent voice and large language model technologies, providing efficient and pre-integrated intelligent interaction solutions for automotive manufacturers around the world. Neusoft will leverage its NAGIC intelligent cockpit platform, combined with Cerence AI's cutting-edge technology, to meet the growing demand for intelligent cockpits in the market.