The Great Shift: How Chinese Open-Source AI Is Overtaking US Models

2/12/2026

The narrative of global artificial intelligence has undergone a profound structural shift. A new analysis by MIT Technology Review dated February 12, 2026, reveals that the past year marked a decisive turning point for Chinese AI. What began with the shock release of DeepSeek’s R1 reasoning model in January 2025 has evolved into a dominant trend where Chinese open-source models are not just competing with Western incumbents but are actively setting the pace for innovation and affordability. The Price-Performance Disruption Just last week, the Chinese firm Moonshot AI released Kimi K2.5, an open-weight model that benchmarks have shown to rival Anthropic’s top-tier Claude Opus. The critical differentiator, however, is not just capability but economics: Kimi K2.5 delivers this frontier-level performance at roughly one-seventh the price of Opus. https://i0.wp.com/asiatimes.com/wp-content/uploads/2024/12/China-AI-Flag-Technology.jpg?fit=1200%2C799&quality=89&ssl=1 This follows the path blazed by DeepSeek R1, which offered reasoning capabilities comparable to OpenAI’s o1 but at a fraction of the cost. The impact was so severe that it triggered a momentary $1 trillion sell-off in US tech stocks. Unlike proprietary US models like ChatGPT, these Chinese models publish their weights, allowing anyone—from researchers to enterprise developers—to download, inspect, and modify the underlying technology. Qwen Overtakes Llama The metrics on developer adoption are clear. On Hugging Face, Alibaba’s Qwen family has overtaken Meta’s Llama models in cumulative downloads. A recent MIT study corroborates this, finding that Chinese open-source models have collectively surpassed US models in total downloads globally. This dominance is fueling a "remix" culture. According to data from the ATOM project, by August 2025, new model variations derived from Qwen accounted for more than 40% of new language-model derivatives on Hugging Face, while Llama’s share fell to about 15%. Qwen has effectively become the default "base model" for the global developer community. Silicon Valley’s New Infrastructure Perhaps the most surprising development is the adoption within the US itself. Martin Casado of Andreessen Horowitz notes that among startups pitching with open-source stacks, there is now an roughly 80% chance they are running on Chinese open models. Data from OpenRouter reflects this surge, showing usage of Chinese models rising from near-zero in late 2024 to nearly 30% in recent weeks. Beyond market share, Chinese labs are driving architectural breakthroughs. DeepSeek’s innovations in model efficiency and memory compression are being widely adopted by Western researchers. As Liu Zhiyuan from Tsinghua University points out, open source has become "politically correct" in the Chinese developer community, serving as a strategic counterweight to US proprietary dominance. In essence, despite export controls and geopolitical friction, Chinese open-source models have transitioned from being mere alternatives to becoming the foundational infrastructure upon which the next generation of global AI is being built.