Tencent's 770B-Parameter H4 AI Model Beats Chinese Rivals in Internal Benchmarks

The Quick Take

• Tencent's new AI model 'H4' outperformed rivals Moonshot AI and Zhipu AI in internal benchmarks.
• The H4 model uses 770 billion parameters and was built on multi-billion-euro investment in AI infrastructure.
• Tencent's AI push intensifies competition among Chinese tech giants racing to match OpenAI's capabilities.

Tencent has released its H4 large language model with 770 billion parameters, reporting that internal benchmarks show it outperforms competing Chinese models from Moonshot AI and Zhipu AI (Z.AI). Handelsblatt's coverage—corroborated across three separate editions of the same report—confirms that Tencent backed the development with multi-billion-euro investment, positioning H4 as its flagship model in the intensifying race among Chinese technology corporations to close the capability gap with OpenAI and Anthropic. The release marks a significant escalation in Tencent's AI strategy, which had previously been more incremental compared to Baidu and Alibaba.

For Chinese technology equities, H4 represents a competitive realignment signal. Tencent's market capitalisation gives it the investment capacity to sustain multi-year model development cycles, putting pressure on smaller Chinese AI startups that lack comparable balance-sheet support. Moonshot AI and Zhipu AI—both venture-backed—face a more difficult fundraising environment if a well-funded incumbent now demonstrably outperforms them. Globally, the H4 release adds to the evidence that the AI competitive moat is narrowing between US and Chinese frontier models, which has implications for US semiconductor export controls and their long-term efficacy in slowing Chinese AI progress.

Watch Tencent's next earnings call for commentary on H4 monetisation—specifically whether the model will be commercialised via cloud API, integrated into WeChat products, or licensed to enterprise customers. The macro variable is US export control policy on advanced AI chips: if restrictions tighten, Tencent's ability to scale H4 training runs becomes the binding constraint, making the current benchmark claims potentially the performance ceiling rather than a baseline. Track Nvidia's sales data to China via third-party channels as an indirect indicator of compute availability for Chinese AI model development.

Synthesized from 3 sources.