Articles

The article provides a detailed interpretation of Zhipu AI's latest released open-source multimodal model GLM-4.5V. The author first showcases GLM-4.5V's excellent performance in multimodal understanding and reasoning tasks, including STEM, spatial reasoning, GUI Agent, and other fields, and emphasizes the significant gains from Reinforcement Learning. Subsequently, the article delves into GLM-4.5V's design concept of being 'reasoning-centric' and conducts a technical analysis from four core dimensions: architecture, pre-training, and post-training (SFT and Reinforcement Learning). In terms of architecture, it emphasizes the design goals of native multimodality, high resolution, and strong temporal understanding, as well as the synergy of the three major components: visual encoder, MLP projection layer, and language decoder. The pre-training section reveals the construction process of high-quality, multi-dimensional data (such as fact-centered paraphrased image-text pairs, interleaved image-text, OCR, and grounding data) and the two-stage long context training paradigm. The post-training elaborates on the key technologies of aligning the model's thinking paradigm (Chain-of-Thought) via SFT and leveraging RLCS (Curriculum Learning Sampling) and a robust reward system to maximize the model's potential, and emphasizes the cross-domain generalization and synergistic effects brought by RL training.

The article provides an in-depth interpretation of Baichuan's latest open-source Medical Large Language Model, Baichuan-M2 (32B), which is licensed under Apache 2.0 and supports commercial use. On the authoritative medical evaluation leaderboard HealthBench, M2 surpasses many closed-source giants and is second only to GPT-5 in terms of medical capabilities. Its core innovations include a Large Verifier System, Mid-Training adaptation to the medical field, and a multi-stage reinforcement learning strategy. This breaks through the limitations of traditional Expert Systems and aims to develop a true AI Doctor, enabling dynamic, multi-round clinical decision-making. The article details how the Verifier System simulates real clinical environments through Patient Simulators and Clinical Scorers, as well as the model's three-stage training architecture in data allocation, reasoning injection, and reinforcement learning. In terms of practical performance, M2 excels in several core scenarios such as emergency referral, medical context understanding, and doctor-patient communication. Especially in the Chinese medical context, it can accurately follow domestic guidelines. In addition, M2 excels in deployment optimization. It compresses the model to within 24GB through Quantization Technology, supporting single RTX4090 GPU operation and compatibility with mainstream Inference Frameworks and Huawei Ascend NPU, which greatly reduces the barrier to entry and cost of localized deployment, especially for medical scenarios with high data security requirements.

As a monthly tech observation report, this article comprehensively reviews the latest developments in global AI for July 2025. The 'Trend Observation' section emphasizes that Chinese LLMs like K2, GLM-4.5, and others have surpassed leading international counterparts in programming, AI Agents, and multi-modal capabilities. Released largely as open-source, these models leverage the open-source ecosystem and cost-effectiveness, solidifying China's central position in the AI competition, suggesting that China and the US are now on par in the language model arena. Simultaneously, the article notes the evolution of image, video, and audio fields towards 'generation-by-understanding,' with 3D generation technology overcoming single-object limitations to enable the creation of combinable parts and complete scenes. AI Coding is advancing towards L4 full automation, while vertical AI Agent applications in finance and imaging are rapidly expanding. The increasing number of mergers and acquisitions suggests a shift in the AI landscape. The industry is transitioning from a period of emerging players (akin to the Spring and Autumn period) to one of intense competition and consolidation (similar to the Warring States period). The 'Time Machine' section meticulously lists key events of the month, including model open-sourcing, application releases, financing, and M&A activities, highlighting the active involvement of Chinese tech giants like Zhipu, Alibaba, and Moonshot AI in open-source AI, alongside updates from international firms such as Hugging Face, Google, and OpenAI, providing readers with a holistic industry overview.

This article details MiniMax's newly launched Agent co-creation platform and the concurrent Agent creation challenge. The platform allows users to create Agents that others can replicate, generating points for the original author. The challenge offers a prize pool of $150,000 in cash and $10,000 in benefits, divided into original and co-creation tracks, encouraging developers to create projects using the MiniMax platform. The author also shares two online games created using MiniMax as examples. The article argues that MiniMax Agent's replication model mirrors the success of the WordPress theme market (like ThemeForest). Templating and replication enable even non-technical users to quickly create applications. This suggests Agent templates could replicate the commercial success of WordPress themes in the AI era, benefiting creators, users, and the platform.

This article details Macaron, an AI Agent developed by Chen Kaijie, which is distinctly different from mainstream, efficiency-focused AI Agents. It focuses on providing emotional value and personalized companionship. Through personal experience, the author demonstrates how Macaron meets users' small needs through gentle interaction, deep understanding of user emotions, and customized gadgets. The article also delves into founder Chen Kaijie's unique background and his insights into human nature, explaining the origin of Macaron's product philosophy of 'understanding you rather than being an obedient tool.' Despite being a seemingly serene product, it is supported by a Deep Memory system trained on an end-to-end hundred-card cluster and stable RL technology based on a 671B large model. The article emphasizes that AI should make life better, not busier, and calls for a human-centered approach to AI.