Search AI Products and News

On June 23, NetEase Youdao officially launched and open-sourced its latest 'Ziyue 3' series large model, named 'Confucius3-Math' in English. This reasoning model, focused on mathematical education, can efficiently run on ordinary consumer-grade GPUs, becoming the first low-cost and high-performance AI educational tool in China. In a series of mathematical reasoning tasks, 'Ziyue 3' has demonstrated outstanding performance that surpasses many large-scale general models. The 'Ziyue Math Model' achieves this through large-scale reinforcement learning and a series of innovative algorithms.

Recently, the ModelScope community announced the release of a new dynamic benchmark dataset named UGMathBench. Its goal is to comprehensively assess the mathematical reasoning abilities of language models across a broad range of undergraduate mathematics subjects. The advent of this dataset fills the current gap in evaluating the reasoning capabilities of language models in the field of undergraduate mathematics and provides researchers with a richer and more challenging testing platform.

The Kimi team recently released a technical report and open-sourced the preview version of Kimina-Prover, including 1.5B and 7B parameter distilled models, the Kimina-Autoformalizer-7B model for data generation, and a revised miniF2F benchmark dataset. Kimina-Prover, jointly developed by the Numina and Kimi teams, is a mathematical theorem proving model that excels in the field of formal theorem proving.

Kunlun Wanwei's Tiangong team announced the launch of the upgraded Skywork-OR1 (Open Reasoner 1) series of models. This follows the February 2025 release of Skywork-o1, the company's first Chinese large language model focused on logical reasoning. The Skywork-OR1 series achieves industry-leading reasoning performance with the same parameter scale, further breaking through the bottlenecks in large model capabilities for logical understanding and complex task solving.

Alibaba is making waves in the AI field with the recent open-source release of its latest multimodal model, Qwen2.5-VL-32B-Instruct. This model is part of the Qwen2.5 series, which also includes 3B, 7B, and 72B versions. The 32B version prioritizes convenient local execution while maintaining performance. Enhanced through reinforcement learning, Qwen2.5-VL-32B excels in several areas. Notably, its responses are more aligned with human expectations.

Recently, the 360 ZhiNao team announced the successful replication of DeepSeek's reinforcement learning results and the official release of the open-source reasoning model Light-R1-14B-DS. This model surpasses DeepSeek-R1-Distill-Llama-70B and DeepSeek-R1-Distill-Qwen-32B in performance, becoming the industry's first 14B parameter model to achieve reinforcement learning effects. It significantly improves mathematical reasoning capabilities, outperforming most 32B-parameter models.

Recently, OpenAI researchers admitted in a newly published paper that despite the advanced nature of current AI technologies, these models still cannot compete with human programmers. OpenAI CEO Sam Altman previously stated that by the end of this year, AI is expected to surpass 'junior' software engineers, but research findings show that these AI models still face significant challenges. Image credit: Image generated by AI, image licensed from service provider Midjourney in the study.

Recently, an AI system developed by Google DeepMind—AlphaGeometry2—has successfully surpassed the average level of gold medalists at the International Mathematics Olympiad (IMO), demonstrating outstanding performance in solving geometry problems. AlphaGeometry2 is an upgraded version of the AlphaGeometry system released by DeepMind last year. According to the latest research from the research team, this system can solve 84% of the geometry problems from the IMO over the past 25 years.

Today, the Alibaba Cloud Tongyi team officially released the new mathematical reasoning process reward model Qwen2.5-Math-PRM. This model offers two sizes, 72B and 7B, with performance significantly outperforming similar open-source process reward models, especially excelling in identifying reasoning errors. The 7B version of Qwen2.5-Math-PRM astonishingly surpasses the widely popular GPT-4o, marking an important step in Alibaba Cloud's research and development of reasoning models.

The Alibaba Qwen team recently published a paper titled 'Lessons Learned from the Development of Process Reward Models in Mathematical Reasoning' and introduced two new models in the Qwen2.5-Math-PRM series, featuring 7B and 72B parameters respectively. These models break through the limitations of the existing PRM framework in mathematical reasoning, significantly improving the accuracy and generalization ability of reasoning models through innovative techniques. Mathematical reasoning has long been a major challenge for large language models (LLMs), especially regarding errors in intermediate reasoning steps.

On the occasion of the annual completion at Beijing Zhihua Huazhang Technology Co., Ltd., the company released the first generation version of its inference model based on extended reinforcement learning technology — GLM-Zero-Preview. This model focuses on enhancing AI's reasoning capabilities, particularly excelling in mathematical logic, code writing, and handling complex problems that require deep reasoning. Compared to the base model, GLM-Zero-Preview shows significant improvements in specialized task capabilities while maintaining general task abilities.