China's Answer to ChatGPT: Alibaba Unveils Trillion-Parameter Model
A price tag akin to an airplane, semi-restricted access, and enormous ambitions.
A price tag akin to an airplane, semi-restricted access, and enormous ambitions.
Chinese giant Alibaba has introduced a new artificial intelligence model, Qwen-3-Max-Preview, boasting over one trillion parameters. It has been made available on the company's official cloud service and on the OpenRouter marketplace. This release continues the Qwen3 series, the first version of which was launched in May with models ranging from 600 million to 235 billion parameters.
Parameters determine the accuracy of a neural network, but the more there are, the higher the energy consumption and computational resource requirements. Expert estimates suggest OpenAI's GPT-4 has between 5 to 7 trillion parameters, remaining one of the largest models globally. Alibaba's new development has surpassed its previous flagship model, the Qwen3-235B-A22B-2507, released in July.
The company published benchmark results: Qwen-3-Max-Preview demonstrated superior capabilities compared to Kimi K2 from MoonShot AI, a simplified version of Claude Opus 4 from Anthropic, and DeepSeek V3.1. The tests covered five different categories, including comprehension of Chinese and English texts, following complex instructions, tackling open-ended tasks, multilingual processing, and tool usage. However, Alibaba has not yet released a full technical report to substantiate these claims. These achievements continue the trend of Qwen 2.5 Max challenging the US monopoly on advanced AI.
The Qwen series has brought Alibaba leadership in the global open-source community: its models have been downloaded over 20 million times on the Hugging Face platform, and the number of derivative solutions exceeds 100,000. This confirms its earlier success when Alibaba's LLM received the highest score in the global Hugging Face ranking. However, Qwen-3-Max-Preview has not been released for open access; it is only available through official channels. The previous model, Qwen2.5-Max, was also not open-sourced. Alibaba engineer Binyuan Hui stated that a version of the model with enhanced "reasoning" capabilities is under development.
The new model is one of the most expensive in the lineup: operational costs are $0.861 per million input tokens and $3.441 per million output tokens. For comparison, the simplified version of Qwen3-235B-A22B-2507 costs $0.287 and $1.147 respectively, while Kimi K2 costs $0.60 and $2.50.
Alibaba also announced massive investments in AI infrastructure—380 billion yuan (approximately $52 billion) over the next three years. This exceeds the company's total investment in this sector over the entire previous decade. This move underscores its ambition to seize leading positions in both the Chinese and global markets.
The company is actively competing with other players in the AI field, regularly introducing new models and open-sourcing some of them. Recent reports have emerged about the development of its own AI processors, aimed at reducing dependence on American Nvidia chips, especially amidst tightening controls by Chinese authorities over foreign technologies. This strategy is becoming particularly relevant as new sanctions against Nvidia could severely restrict Chinese companies' access to advanced chips. It is also worth considering the potential risks associated with Chinese tech companies, as warned by Western intelligence agencies in the context of the "Chinese shadow over Europe." The release of Qwen-3-Max-Preview has solidified Alibaba's status as a serious competitor in the large language model segment and confirmed China's intent to influence the future development of the industry.
