Conceptual diagram of AI Photo: VCG

Chinese AI powerhouse DeepSeek on December 1 released two major upgrades - DeepSeek-V3.2 and the high-compute DeepSeek-V3.2-Speciale - continuing its aggressive campaign to be shoulder-to-shoulder with its global rivals.

The company noted that DeepSeek V3.2 combines stronger reasoning, efficient tool use, and a sparse-attention design that drastically cuts costs for ultra-long contexts, while the Speciale version targets advanced mathematics and long-horizon planning.

The announcement comes amid a heated global AI sprint. OpenAI launched GPT-5 version in August as its smartest model yet, and Google rolled out Gemini-3.0-Pro in November. Competition now spans raw intelligence, latency, cost, tool-chaining fluency, and real-world reliability. Winners will be those who deliver elite capability by a massive scale while at a relatively lower cost.

Industry analysts said that Chinese labs are rewriting release cycles. Unlike overseas giants that unveil flagship models every 6 to 12 months, Chinese stars like DeepSeek, Moonshot AI, and Alibaba now ship production-grade versions weekly or biweekly, often in multiple variants. This "high-frequency + open weights" model is helping them win growing user adoption in the world.

Trend of agentic AI

DeepSeek said that its V3.2 matches GPT-5 across key compute metrics through intensive post-training reinforcement learning, while maintaining far higher operational efficiency. It is also the first DeepSeek model to natively fuse thinking and tool-use, supporting both reasoning-first and direct-action modes. 

Its Speciale variant reportedly exceeds GPT-5 and approaches Gemini-3.0-Pro in complex reasoning, scoring at gold-medal level on 2025 International Mathematical Olympiad and International Olympiad in resolving Informatics problems.

On November 6, Moonshot AI released Kimi K2 Thinking, an agent that reasons step-by-step while chaining up to 200-300 tool calls autonomously. It posted major gains in coding, writing, agentic search, and general capabilities, showcasing the rapid convergence of deep reasoning and tool execution system.

Tian Feng, president of the Fast Think Institute and former dean of Chinese AI software giant SenseTime's Intelligence Industry Research Institute, told the Global Times that future AI agents will merge long-memory planning with seamless service execution.

Unlike early agents limited to search, new Chinese AI agents already integrate shopping, payments, logistics, social, and entertainment - becoming true "all-purpose butlers." Tian predicted that specialized vertical agents in law, healthcare, manufacturing, and government services will be welcomed by users, as it is orchestrated by general-purpose agents in a collaborative ecosystem.

Marching globally

Global validation continues to poor in. On November 26, the Financial Times cited a study by the Massachusetts Institute of Technology and Hugging Face showing that total share of downloads of new Chinese-made open models rose to 17 percent in the past year. The figure surpasses the 15.8 percent share of downloads from American developers, said the report.

China's push to release open-source AI models comes in stark contrast to the "closed" approach of most of the biggest US tech companies, the report noted.

Alibaba accelerated the consumer push on November 24 with the public test of its Qwen App, which garnered more than 10 million downloads within the first week of its launch, outpacing ChatGPT, Sora, and DeepSeek to become the fastest-growing AI model application to date, according to Securities Times.

The crucial reasons behind Chinese AI models' download explosion are evident, and their performance now matches those closed AI models, combined with fully open weights and dramatically lower deployment costs, Liu Gang, chief economist at the Chinese Institute of New Generation AI Development Strategies, told the Global Times on Thursday.

Even overseas institutions are switching. AI Singapore (AISG) is undergoing a major strategic shift, moving away from Meta's model to adopt Alibaba's open-source Qwen architecture for its latest Southeast Asian language model project. The decision marks a notable expansion of China's open-source AI footprint on the global stage, according to information Alibaba shared with the Global Times.

China's AI industry size exceeded 700 billion yuan in 2024, maintaining a growth rate of more than 20 percent for  the past five consecutive years, according to a report released by China Internet Network Information Center in July. The report noted that in the first half of 2025, generative AI models made comprehensive progress from technology to application, with the number of new models keeping on surging and use scenarios continuing to expand.

Looking ahead, Liu said that pure text large language models have hit diminishing returns; genuine intelligence demands a fused understanding of vision, audio, and physics, adding that vision-dominant world models for autonomous driving and robotics are advancing at the fastest pace.

China is seizing pole position in this decisive arena, powered by the world's largest multimodal datasets including short video, industrial vision, unmatched edge-deployment optimization, and an uncompromising open-source philosophy that lets the global developer community amplify every breakthrough, industry analysts said.