Liu Wei, former head of Tencent Holdings' Hunyuan foundational model team, says China is falling behind in the global large language model race — but argues the country can still carve out a winning position in AI through a different strategic path. Speaking publicly for the first time since his departure from Tencent in late 2024, Liu offered a candid assessment of where Chinese AI stands and where it must go.

The Paradigm Problem

Liu Wei, who spent more than eight years at the Shenzhen-based tech giant before quietly exiting, said the central weakness of China's AI industry is a lack of fanshi — a Chinese term for 'paradigm' widely used among AI researchers to describe a breakthrough that defines a new era of innovation. Notable examples of such paradigm shifts, he noted, include OpenAI's ChatGPT and Anthropic's Claude Code. 'Chinese companies are either copying DeepSeek or US companies at the core technical level,' he said, referring specifically to large language model development.

Benchmark Scores vs Real-World Performance

There has been recurring speculation since DeepSeek's breakout moment in early 2025 about whether Chinese LLMs have genuinely closed the gap with their US counterparts. Liu cautioned that narrowing scores on public benchmarks do not accurately reflect the gap in real-world usefulness — a distinction that matters significantly for enterprise and consumer adoption. The divergence between benchmark performance and practical capability remains one of the most debated fault lines in global AI evaluation.

Why It Matters

The timing of Liu's departure raised immediate questions: Hunyuan had been introduced only a year before he left, and Tencent remains one of China's most heavily capitalised technology companies. His exit suggested internal friction over strategic direction at a critical juncture in the AI arms race. His assessment now lends weight to concerns that China's LLM ecosystem is increasingly derivative rather than generative at the frontier level.

The Competitive Backdrop

US industry leaders have continued pushing the technical frontier, most notably with the launch of Anthropic's Mythos model in April. Meanwhile, domestic Chinese leader DeepSeek reportedly failed to match the heights of its earlier breakthroughs with its latest V4 model, according to industry observers. Other major Chinese tech players — including ByteDance, Kuaishou, and Meta Platforms' China-facing operations — continue to invest heavily in LLM development, but the paradigm gap Liu describes remains unaddressed.

What's Next

Despite his pessimism on LLMs, Liu Wei maintains that China retains a viable path to winning in AI more broadly — likely through application-layer innovation, vertical deployment, and markets where access to frontier US models is restricted. The question of whether any Chinese lab can produce a genuine paradigm-defining breakthrough, rather than an optimisation of existing architectures, will define the next phase of the global AI competition. Observers will be watching DeepSeek's next model release and Tencent Hunyuan's roadmap for early signals.