China's AI Milestones in 2025: DeepSeek, Qwen, and the Open-Source Revolution

Milestones of China in AI of 2025

TLDR: China overtook the US in open-source AI model downloads in 2025, with DeepSeek's $6M frontier model shattering cost assumptions and Alibaba's Qwen becoming the world's largest open-source model ecosystem. Meanwhile, domestic AI chips gained ground, AI agents went mainstream, and talent started flowing back from Silicon Valley.

This deep dive from AI Supremacy, featuring Tony Peng's analysis, catalogs an extraordinary year for Chinese AI that upended multiple Western assumptions about what's required to build frontier AI systems.

The DeepSeek moment in January 2025 stands out as the year's defining event. Nine days before Chinese New Year, this previously unknown lab released DeepSeek-R1, an open-weight model matching OpenAI's o1 in reasoning benchmarks. The shocking detail: the base model cost under $6 million to train, and R1 itself under $300,000. For inference, DeepSeek charges $0.55 per million input tokens versus OpenAI's $15.00. Without any marketing, the app shot to the top of iOS charts in both the US and China, triggering a significant US tech stock selloff as investors questioned their assumptions.

The open-source numbers tell a broader story. Chinese models now account for 17.1% of global downloads, surpassing the US for first place. Alibaba's Qwen family accumulated over 400 million downloads and spawned 140,000 derivative models. Their latest Qwen3-Max with one trillion parameters outperforms Claude Opus 4 and DeepSeek V3.1 on agent benchmarks. Perhaps most telling: Cursor and Cognition, two of Silicon Valley's hottest startups, are reportedly building on top of these Chinese open models.

The agentic AI story gained substance with Manus, a Beijing startup that launched an autonomous agent capable of browsing the web, analyzing data, and executing complex tasks. It outperformed OpenAI's Deep Research on GAIA benchmarks and raised $75 million from Benchmark Capital at a $500 million valuation within a month of launch. By December, Manus claimed $100 million ARR.

On the hardware front, Huawei went public with its chip roadmap after years of secrecy. Their next-generation AI supercluster is projected to reach 524 EFLOPS by 2026, potentially one of the world's most powerful AI computers. Domestic GPU companies Moore Threads and MetaX went public with 425% and 693% first-day gains respectively. The twist: late 2025 saw the Trump administration allow Nvidia H200 sales to China, creating a dilemma — Beijing can't simply abandon domestic chip development, but Nvidia GPUs remain the best option for pre-training frontier models.

ByteDance's Doubao chatbot dominated with 160 million monthly active users, and in December they partnered with ZTE to create what may be the first truly smart phone — an AI assistant that can independently control your device, see the screen, understand interfaces, and complete tasks across apps. WeChat, Alipay, and other major apps immediately blocked it, revealing how threatened the existing app ecosystem feels.

The robotics spectacle was remarkable — Xpeng's humanoid IRON walked with such fluid grace that people insisted it was a person in costume. Unitree's G1 robots performed standing side flips. Yet the honest assessment: software brains that would make these machines genuinely useful remain far away, limited by the lack of diverse physical training data.

Perhaps the most strategically significant development: talent flow is shifting. Google DeepMind's VP of Research joined ByteDance. An OpenAI researcher became Tencent's Chief AI Scientist. While the US still retains 87% of Chinese AI researchers, it's attracting fewer new ones, and China's maturing industry now offers competitive salaries and leadership positions.

For architects and teams watching global AI development, the implications are substantial. The cost assumptions behind frontier AI are being challenged. Open-source is proving a viable path to competitiveness. And the infrastructure race — chips, data centers, energy — is becoming as important as the algorithm race.

Key takeaways:

• DeepSeek-R1 shattered the "hundreds of millions required" narrative for frontier models
• Chinese open-source models now power nearly 30% of global AI usage, up from 1.2% in late 2024
• AI agents moved from demos to products with real revenue (Manus: $100M ARR)
• Domestic AI chips are gaining market share, with Huawei Ascend at 35-40% of Chinese market
• Talent flow from Silicon Valley to China, while small, would have been unimaginable two years ago

Tradeoffs:

• Open-weight models gain global adoption but expose innovations to competitors
• Domestic chips enable sovereignty but lag Nvidia in performance and ecosystem support
• Massive AI app ecosystems drive adoption but create moats blocking new entrants like phone AI agents

Link: Milestones of China in AI of 2025

This article was compiled from the Substack newsletter. The opinions and summaries presented are interpretations of the original sources — always read the linked articles for complete context.