Zhipu has officially released the GLM-4.7 large language model, focusing on real-world development environments. It significantly improves multi-step task handling, complex reasoning, and tool-calling capabilities, supporting multilingual programming and terminal agent scenarios. It achieves excellent performance on mainstream code and tool benchmarks such as BrowseComp and τ²-Bench, becoming the top Chinese self-developed and open-source model on platforms like Code Arena. GLM-4.7 has become the default model for the GLM Coding Plan and is being promoted through BigModel.cn API and multi-platform integrations to drive the engineering application of the open-source ecosystem. The company's revenue CAGR from 2022 to 2024 reached 130%, with revenue of 190 million RMB in the first half of 2025, primarily driven by related large model business.

NVIDIA has released NitroGen, an open-source vision-action foundation model. Trained on over 40,000 hours of internet videos from more than 1,000 games, it learns directly from pixels and gamepad operations. It features a unified control space, a universal simulator, and multi-game benchmarks. Utilizing a Diffusion Transformer architecture, it achieves an average 45%-60% completion rate across multiple task types in zero-shot settings. Performance on new games improves by 10%-52% after fine-tuning, laying the foundation for generalist game AI and multi-task agents.

Liquid AI has introduced the LFM2-2.6B-Exp experimental model. Based on a 2.6B parameter, 32K context efficient hybrid architecture, it employs pure Reinforcement Learning (RL) fine-tuning, focusing on instruction following, knowledge, and mathematical tasks. It surpasses large-scale models like DeepSeek R1-0528 on benchmarks such as IFBench. It supports multilingual capabilities, native tool calling, and dynamic hybrid reasoning, making it suitable for edge devices and Agent systems. The weights have been open-sourced.

DeepSeek has released Math-V2, a 685B parameter model focused on mathematical reasoning and theorem proving. It has achieved gold medal and even super-human performance in top competitions like the IMO, CMO, and Putnam 2024. It scores nearly 99% on the base set of IMO-ProofBench, significantly leading closed-source models like Gemini DeepThink. The model innovatively employs 'meta-verification' multi-level supervision, forcing the AI to demonstrate rigorous step-by-step reasoning and eliminating 'gaming' behavior that only seeks final answers, significantly enhancing the reliability of reasoning chains. The model and its weights are fully open-sourced, driving breakthroughs in the open-source ecosystem for high-level reasoning.

The AI industry in 2025 shows three major trends: First, open-source large models (e.g., GLM-4.7, DeepSeek series) continue to break through in areas like reasoning, coding, and mathematics, challenging the dominance of closed-source giants. Second, Agent intelligence is moving from experimentation to practical production, with a surge in enterprise deployments driving automation and business restructuring. Third, AI engineering, compute infrastructure, and data governance have become key to implementation. AI's deep integration with industries like finance, manufacturing, and healthcare is driving a new round of industrial upgrades.

China's AI chip industry is moving towards a 'Huawei Ascend + Multiple Strong Players' landscape, with domestic AI chip self-sufficiency expected to exceed 70% by 2026. Leading companies are accelerating their procurement of domestic computing power. Open-source large model providers like Zhipu and MiniMax are enhancing the international competitiveness of Chinese AI models through ecosystem integration and global open-source community collaboration. Hong Kong is becoming a new hub for AI company listings and international cooperation.

In 2025, AI Agents transitioned from experiments to production. Companies like LTIMindtree and Coforge achieved 70-90% business automation through multi-level Agent systems. Alibaba's DingTalk launched Agent OS, supporting unified management and collaboration of multiple agents. Platforms like UiPath are driving Agent orchestration and security governance. Agents have become a core tool for enterprise cost reduction, efficiency improvement, and process re-engineering.

NVIDIA is leading the global shift in data centers from AC to 800V DC power to support next-generation high-power AI systems, with liquid cooling becoming mainstream, forcing an industry-wide upgrade. The compute infrastructure upgrade benefits upstream and downstream companies including domestic chips and optical modules. AI infrastructure has become a new high ground in global technology competition.

In 2025, content companies and AI giants engaged in multiple rounds of litigation over copyright for AI training data. Licensing agreements between industries like news, music, film/TV and AI companies have become a new trend. China introduced management measures for AI personification interactive services, emphasizing safety, reliability, and value orientation, promoting compliant governance and healthy industry development for AI content.

Multimodal large models like Qwen3 VL 32B integrate vision, language, and reasoning, supporting million-level context windows, and possess capabilities such as 2D/3D object localization, code generation, and long-video understanding. They lead mainstream models like Gemini 2.5 Pro on various perception and STEM reasoning benchmarks, driving new applications for multimodal agents and digital interface operations.