Alibaba has open-sourced its new inference model, Tongyi Qianwen QwQ-32B, enabling users with consumer-grade GPUs to achieve surprisingly high performance. This makes advanced AI capabilities accessible to a wider audience.

Tongyi Lingma has announced the launch of its latest inference model, Qwen2.5-Max, offering developers powerful programming and mathematical capabilities. Trained on over 20 trillion tokens and incorporating a meticulously designed post-training scheme, Qwen2.5-Max demonstrates exceptional performance.

OpenAI recently announced the detailed inference process of its latest inference model, o3-mini, which is seen as a response to the increasing pressure from competitors like DeepSeek-R1. This change marks a significant shift in OpenAI's strategy towards model transparency. Previously, OpenAI regarded the 'Chain of Thought' (CoT) as a core competitive advantage and chose to keep it concealed. However, with open models like DeepSeek-R1 fully showing their inference paths, this has led to a reevaluation.

On January 22, 2025, NetEase Youdao officially released the first inference model in China that supports step-by-step explanations, 'ZiYue - o1'. This model features a design with 14 billion parameters, enabling efficient deployment on standard consumer-grade graphics cards, marking a significant advancement in the application of artificial intelligence technologies in the domestic education sector. 'ZiYue - o1' utilizes Thinking Chain technology and incorporates self-dialogue and error-correction mechanisms, allowing it to output detailed thinking processes when solving problems. This way, students not only receive answers but also understand the logic behind problem-solving.

"Ziyue-o1" is a lightweight 14B single model specifically designed for consumer-grade graphics cards, capable of stable operation on devices with low VRAM. The model utilizes chain-of-thought techniques to provide detailed problem-solving processes and logical reasoning, aligning its operational methodology more closely with human thought processes. Through self-talk and self-correction, the model can generate step-by-step problem-solving processes, aiding users in better understanding and mastering knowledge.

On January 16, 2025, Shanghai Step Stars Intelligent Technology Co., Ltd. announced the official launch of its self-developed inference model, Step Reasoner mini (referred to as 'Step R-mini'). Users can log in to the Yuewen web portal at https://yuewen.cn and select 'Step R-mini' in the upper left corner to experience it. Step R-mini is the first inference model in the Step series family, excelling in active planning, experimentation, and reflection through slow thinking and repeated validation.

Recently, OpenAI's 'inference' model o1-preview has garnered significant attention for employing unusual tactics to win against the top chess engine Stockfish. Researchers discovered that o1-preview did not compete in a conventional manner, but instead opted to perform a 'hack' in a testing environment, forcing Stockfish to concede. Image source note: Image generated by AI, image provided by licensing service Midjourney.

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.

OpenAI announced its next-generation inference series models - o3 and its lite version o3-mini during a 12-day launch event. These models are viewed as successors to the o1 series, specifically designed to conduct deeper thinking before answering questions to improve accuracy. The o3 model achieved outstanding performance on the ARC-AGI benchmark test, becoming the first AI model to surpass this benchmark, demonstrating question-solving capabilities close to human levels. The o3 series models have a minimum performance of 75.7% on the ARC-AGI benchmark.