Meta has acquired Singapore-based AI Agent company Manus for over $2 billion. Manus, which has garnered attention in Silicon Valley for its general-purpose agent products and over $100 million in annualized revenue, saw this acquisition not only mark a shift in the AI landscape from model competition to application deployment and commercialization but also signal an impending wave of mergers and acquisitions in the Agent field. The Manus team will maintain independent operations, and its AI Agent capabilities will be integrated into Meta's products like Facebook, Instagram, and WhatsApp. The deal has drawn regulatory scrutiny in both China and the US due to Manus's Chinese background, with Meta pledging to sever ties with Chinese investors and exit the Chinese market post-acquisition. Manus, previously backed by Benchmark, Tencent, ZhenFund, etc., had a pre-IPO valuation of $2 billion. This M&A reflects the AI industry's high regard for Agent products with actual revenue-generating capabilities and highlights the influence of Chinese AI startups in the global market.

Tencent's HY-MT 1.5 translation model has been open-sourced, comprising 1.8B and 7B versions supporting 33 languages and 5 dialects. After quantization, the 1.8B model requires only 1GB of memory to run offline on terminals like mobile phones, with performance on authoritative benchmarks like FLORES-200 approaching that of closed-source large models such as Gemini-3.0-Pro. The model supports features like custom terminology libraries, contextual understanding, and format preservation. It utilizes On-Policy Distillation technology where a large model guides a small model, significantly enhancing the small model's reasoning and translation capabilities.

From late 2025 to early 2026, Chinese AI companies like MiniMax (Xiyu Tech), Zhipu AI, and Insilico Medicine have intensively launched IPOs on the Hong Kong Stock Exchange. MiniMax, founded less than three years ago with an average team age of 29, has over 210 million global users. Revenue for the first nine months of 2025 grew 170% year-on-year, with overseas revenue exceeding 70%. The IPO could value the company at over HKD 50 billion, with 14 cornerstone investors including Alibaba and Eastspring subscribing. Zhipu AI, known as one of the 'AI Six Tigers', is listing under Chapter 18C, with its GLM-4.7 model excelling in areas like programming. Insilico Medicine, as an AI drug discovery representative, is raising HKD 2.3 billion, with participation from international giants like Lilly and Temasek. Its AI drug discovery platform Pharma.AI has developed over 20 clinical assets. Chinese AI firms, with their self-developed multimodal large models, global product deployment, and high growth potential, have gained strong recognition from international capital and industrial giants, emerging as a new force in the global AI industry.

In 2025, the large language model industry continued its rapid development, with reasoning capabilities (RLVR), multimodality (text, speech, vision), vertical domain specialization, and intelligent agent (Agent) applications becoming the main themes. DeepSeek emerged as a dark horse with its open-source strategy and reasoning capabilities, with new models like V3.2 competing with closed-source models like GPT-4o on tasks such as math and coding. Leading domestic players like Tencent, Alibaba, Baidu, ByteDance, and Kimi made continuous breakthroughs in model capabilities, application ecosystems, and user scale. Technically, innovations like MoE architecture, dynamic sparse activation, multimodal fusion, end-to-end speech models, and unified Transformer architectures were frequent. At the application level, vertical industry agents, localized Agents, ambient programming, visual interaction, and privacy protection became new trends. The industry faces challenges like data diversity, model interpretability, and standards. The future will focus on dual drivers of fundamental capability enhancement and scenario deployment.

In mid-2025, Menlo research showed that Anthropic's Claude model held a 32% share in enterprise LLM production environments, surpassing OpenAI (25%) and Google Gemini (20%), with Meta Llama at 9% and DeepSeek at 1%. Enterprise LLM API spending doubled in six months to $8.4 billion, primarily driven by Claude 3.5/3.7's excellent performance in scenarios like code generation. Closed-source models dominate enterprise procurement due to compliance and security advantages, with developers often adopting multi-model strategies to avoid vendor lock-in. While the report sample is limited, it reflects a strong enterprise market demand for high-performance, controllable LLMs. It speculates that inference costs will dominate budgets in 2026, long-cycle Agent applications will become mainstream, and enterprises need to continuously monitor model performance, cost, and governance capabilities.

As AI applications proliferate, global data center energy consumption and carbon emissions have risen significantly. The University of California, Riverside proposed a 'Federated Carbon Intelligent (FCI)' system that intelligently schedules AI tasks by monitoring server health and grid carbon intensity in real-time, balancing emission reduction with hardware lifespan extension. Simulation results show FCI could reduce carbon emissions by 45% over five years and extend server lifespan by 1.6 years. The solution requires no new hardware, emphasizes hardware-software synergy and full lifecycle carbon management, providing a new approach for greening AI infrastructure. The research team is collaborating with cloud service providers to implement FCI in actual data centers, helping the global AI industry achieve NetZero goals.

The U Lab team at the University of Illinois Urbana-Champaign has released the LLMRouter open-source routing library, supporting dynamic selection of the optimal model from a multi-LLM pool based on task complexity, quality, and cost. The system includes over 16 built-in routing algorithms covering single-turn, multi-turn, personalized, and Agentic routing, supporting various strategies like KNN, SVM, Elo, and Graph Neural Networks. LLMRouter provides a complete data generation pipeline covering 11 mainstream benchmarks, supports plugin extensions and a Gradio interactive interface. Personalized routing learns user preferences via heterogeneous graph learning, and Agentic routing supports multi-step reasoning and complex task distribution. This system provides infrastructure for enterprises and developers to achieve efficient, controllable, multi-model collaborative LLM inference, driving AI inference services towards intelligent scheduling and personalization.

In 2025, AI industry investment and M&A were active, with large deals like Meta acquiring Manus, Salesforce acquiring Informatica, and HPE acquiring Juniper occurring frequently. Nvidia's market cap surpassed $5 trillion, maintaining its dominance in the AI chip market, and it reached a $500 million investment and joint development agreement with Intel. AI data centers and AI PCs became new growth points, with giants like Microsoft, Amazon, and Google increasing infrastructure investments. AI Agent and Agent Orchestration protocols (like MCP, A2A) drove improvements in AI system interoperability. Debates over an AI bubble intensified, with some enterprises showing lower-than-expected returns on AI investment, and attention focused on compute leasing cycles and capital structures. Overall, the AI industry is shifting from model innovation to application deployment, infrastructure upgrades, and ecosystem consolidation.

In 2025, technology and applications in the large model field saw comprehensive upgrades. Reasoning capabilities (RLVR+GRPO) became mainstream, with open-source models like DeepSeek R1 achieving reasoning and explanation capabilities comparable to closed-source models at lower cost. Efficient fine-tuning methods like LoRA and DPO became widespread, while new architectures like MoE, sparse attention, and Mamba improved inference efficiency. At the toolchain level, Agentic AI, the MCP protocol, spec-driven development, and inference-based evaluation systems (like LLM-Judge, IRT) became new hotspots. Industry focus shifted from single model performance to reasoning chains, tool calling, inference token costs, and context engineering. Application scenarios like AI programming, AI Agent, AI PC, and data centers rapidly deployed. Industry evaluation systems became more diversified and practical. Future trends include scaling reasoning models, optimizing inference tokens, continuous learning, industry-specific models, and local deployment.