Chinese AI models gain favor overseas

Alibaba Group has launched its 2.4-trillion-parameter Qwen3.8-Max model that can analyze 200-page financial reports and process more than 100 hours of video content, putting it directly among the world's top-tier large language models.

Several sources told China Daily that Alibaba's latest Qwen model has entered the "deep testing" phase in Tesla's vehicle systems in China and is expected to be used for Tesla's in-car platform soon.

The development comes as more and more companies from the United States adopt Chinese artificial intelligence models, which they believe offer frontier-level performance at substantially lower costs.

In the latest Arena benchmark rankings, one of the most widely followed third-party evaluations of AI model performance, Alibaba's Qwen models ranked just behind US-based Anthropic's Claude series.

Alibaba said that Qwen3.8-Max is expected to become open-source next week, following the release of Qwen3.8-27B, further expanding access for developers building AI applications.

Qwen3.8-Max's application programming interface pricing in China has been set at 12 yuan ($1.67) per million input tokens and 36 yuan per million output tokens, with cached input tokens priced at 1.5 yuan per million tokens.

International pricing is also significantly lower than comparable frontier models, with input and output costs respectively at about 40 percent and 24 percent of Anthropic's Claude Opus prices.

Chinese AI models have been continuously gaining traction overseas. According to OpenRouter data, weekly usage of Chinese LLMs stood at 14.1 times that of US models as of July 27, marking the 13th consecutive week that Chinese models ranked first on the platform in terms of weekly token consumption.

Late last month, Coinbase CEO Brian Armstrong said the cryptocurrency company had adopted China-developed AI models, including Zhipu AI's GLM-5.2 and Moonshot AI's Kimi K2.7, as default tools for engineers through an internal AI gateway. Through model switching, intelligent routing and caching optimization, Coinbase has cut its AI costs by nearly half.

San Francisco-based AI startup Lindy switched from Claude to DeepSeek-V4 in June, and said its API expenses had exceeded its entire payroll costs before the transition.

"It is not about low pricing anymore," said Wang Peng, a researcher at the Beijing Academy of Social Sciences. "Companies globally are increasingly weighing performance against cost, privacy, flexibility and suitability for AI agents and enterprise workflows."

Morgan Stanley said in a research note that Moonshot AI's Kimi K3, the full model weights of which became officially open-source on July 27, represented a broader trend of Chinese LLMs catching up with US leaders in model scale, performance and pricing.

Peng Jiahao, a researcher at the Shanghai Artificial Intelligence Research Institute, said: "The pace of AI development is increasingly moving in China's favor. But both Chinese and US AI models are still facing challenges in reaching short-term breakeven and the AI industry is still far from achieving widespread profitability."