TLDR: In 2025, Chinese AI agents like Tencent’s Hunyuan, Monica’s Manis, Alibaba’s Moonshot AI, MiniMax M1, and ByteDance’s Seedance 1.0 have made significant advancements, showcasing capabilities from 3D world creation and autonomous task completion to massive context windows and cinematic video generation. These developments highlight China’s rapid progress in the global AI race, challenging the dominance of Western models and fostering an open-source ecosystem.
The year 2025 has witnessed a remarkable surge in the capabilities and global presence of Chinese artificial intelligence agents, prompting discussions about their long-term impact on the AI landscape.
While the United States has historically led in AI development, new reports and product launches indicate China is rapidly closing the gap, particularly in the quality and quantity of its AI models.
Among the standout innovations is Tencent’s Hunyuan, which recently unveiled Hunyuan World 1.0, capable of generating entire 3D worlds from simple text descriptions. Complementing this, Huan 3D 2.5 creates highly detailed 3D objects with 1024 resolution geometry, a four-fold improvement over previous versions. These models are notable for their open-source nature and accessibility, running efficiently on mobile phones and laptops, and boasting an impressive 256k context window, allowing them to process vast amounts of information.
Another significant player is Monica’s Manis, an autonomous AI agent launched in March 2025. Within just one week of its release, Manis garnered over two million users on its waitlist. What sets Manis apart is its ability to not just chat, but to actively perform tasks by navigating websites, comparing options, and completing actions like bookings, leveraging a ‘virtual computer‘ interface. Manis intelligently integrates multiple AI models, including Anthropic’s Claude and Alibaba’s Quen, to select the most suitable model for each specific task, demonstrating a ‘super agent‘ approach to AI.
Alibaba-backed Moonshot AI, based in Beijing, is ambitiously pursuing Artificial General Intelligence (AGI). Its flagship chatbot, Kimmy, can process over 200,000 Chinese characters in a single conversation, equivalent to analyzing entire books or large codebases without losing coherence. Moonshot’s architecture emphasizes long-context understanding, multimodal learning, and self-improvement, enabling its models to reason, learn, and adapt autonomously.
In the realm of large language models, MiniMax M1, released in mid-2025, stands out as an open-source mixture-of-experts model. It boasts an extraordinary 1 million-token context window, allowing it to analyze extensive documents like entire book series or court transcripts in one go. Furthermore, M1 can generate up to 80,000 tokens in a single response, akin to a full novel chapter or a detailed research paper. Its innovative lightning attention system also ensures faster and more efficient processing. Remarkably, the entire training cost for M1 was reported to be under $600,000, challenging the notion that top-tier AI models require billions in investment.
ByteDance’s Seedance 1.0 is revolutionizing video generation. This autonomous AI video model allows users to create multi-shot cinematic videos from text or images, producing 1080p, 24fps sequences with lifelike motion and smooth camera transitions.
These advancements underscore China’s strategic push for AI self-sufficiency, particularly in response to U.S. export controls on advanced chips and technologies. While Chinese chips may still lag behind American counterparts in some aspects, companies like Huawei are narrowing the gap by clustering more chips to boost performance. China has also advocated for an international open-source community for AI model deployment and improvement, a vision that appears to be gaining traction in the Global South.
Also Read:
- Global AI Agent Competition Ignites Innovation and Participation for 2025 Challenge
- China Unveils ‘Agent Hospital’: A Virtual Healthcare Revolution Powered by AI
Experts note that Chinese models have rapidly improved in recent years, significantly narrowing performance gaps with U.S. systems in language understanding and general reasoning tests. For instance, in advanced language tests (MMLU-Pro), Chinese platform DeepSeek-R1 has even surpassed American-made models. This rapid progress suggests that while the U.S. remains a leader, China is poised to become a dominant force in the 21st-century AI landscape, demonstrating that innovation can emerge from anywhere, benefiting the global AI community.
