Skip to content 
Alibaba Group Holding has launched its biggest artificial intelligence model so far, and it is taking part in a trillion-parameter race with other competitors worldwide. The firm posted the launch of Qwen-3-Max-Preview, putting itself at the same level as OpenAI and Google DeepMind in the development of the best systems. Executives indicated that the model makes significant text comprehension, instruction handling, and multilingual improvements.
Alibaba's breakthrough in trillion-parameter AI systems
The company claims that Qwen-3-Max-Preview is the first in the company portfolio to exceed one trillion parameters. Parameters are those variables that determine how an AI system works with information, and bigger models would typically offer better performance at the cost of large computing resources. The new launch is based on the Qwen3 series that was launched in May with models that ranged between 600 million to 235 billion parameters.
Alibaba Cloud has provided the model on its official platform and the third-party market OpenRouter. It was internally tested as a step forward compared to the high-end system used previously by the company, the Qwen3-235B-A22B-2507, which was launched in July. Benchmark scores published by Alibaba indicated that Qwen-3-Max-Preview was the best at five metrics out of competition with other rival models: the Kimi K2 of MoonShot AI, the Claude Opus 4 non-reasoning version, and DeepSeek V3.1. These results, however, were not followed by a formal technical report.
Performance gains and access details
Alibaba stated that Qwen-3-Max-Preview shows substantial progress in overall ability. Enhancements include Chinese-English text understanding, complex instruction following, tool invocation, and the ability to manage open-ended tasks. Engineers involved in development said that a “thinking” version of the model is currently being prepared for release. Unlike earlier models in the Qwen3 series that were open-sourced, the new trillion-parameter version has not been released publicly.
Access is restricted to official channels, reflecting Alibaba’s strategy with previous Max series systems such as Qwen2.5-Max.The company has implemented a tiered pricing structure. Charges begin at US$0.861 per million input tokens and US$3.441 per million output tokens. This makes the model one of the costliest options in the Qwen portfolio. By comparison, the Qwen3-235B-A22B-2507 is priced at US$0.287 for input and US$1.147 for output per million tokens, while Kimi K2 costs US$0.60 for input and US$2.50 for output.
Expanding investments and ecosystem growth
Alibaba has invested in AI infrastructure more than 380 billion yuan, or about 52 billion US dollars in the next three years. According to executives, this commitment is higher than the amounts the company has spent on the sector in the last ten years. Financial reports showed that products related to AI have recorded triple-digit growth in eight quarters, and this is a pointer that the investment strategy is paying off. The Qwen models have enhanced the strength of Alibaba in the open-source ecosystem globally. The statistics of the company revealed that previous Qwen releases have already obtained over 20 million downloads and have motivated 100,000 derivative models on the platform Hugging Face, where they are built.
The firm emphasized that its model architecture scaling has resulted in performance improvements that are quantifiable and also stated that the ultimate Qwen-3-Max version would bring additional improvements. According to industry analysts, the GPT-4.5 version of OpenAI is also thought to have between 5 trillion and 7 trillion parameters, which may imply that the trillion-parameter system offered by Alibaba continues to rank below the largest competitors. However, the introduction shows that the company in Hangzhou has ambitions to compete in the development of higher-order models as it increases its presence in enterprise and consumer AI applications.
