Alibaba challenges OpenAI

The Qwen-3-Max-Preview model has over a trillion parameters. Photo: Reuters .

Alibaba Group Holding has just released its largest artificial intelligence model to date, marking a new milestone in the global AI race. Called Qwen-3-Max-Preview, this is the first model from the Chinese tech giant to surpass one trillion parameters, placing Alibaba among the giants like OpenAI, Google DeepMind, and Anthropic.

The new model, launched on September 5th, is now available on the Alibaba Cloud cloud service platform and the large OpenRouter modeling language marketplace. Qwen-3-Max-Preview continues the Qwen3 series, which was first introduced in May with versions ranging from 600 million to 235 billion parameters.

In the field of AI, parameters are considered the "intelligence" of a system, which are variables that are adjusted during the training process. A larger number of parameters generally means greater capabilities and also requires enormous computing power. For comparison, OpenAI's GPT-4.5, one of the largest models currently available, is estimated to have between 5 and 7 trillion parameters.

According to the announcement, Qwen-3-Max-Preview is a text processing model that outperforms Qwen3-235B-A22B-2507. In internal tests, Qwen-3-Max-Preview achieved higher results than MoonShot AI's Kimi K2, a non-inference-capable version of Claude Opus 4, and DeepSeek V3.1 which supports Anthropic formatting. However, these benchmarks have not yet been included in the official technical report.

“Qwen-3-Max-Preview shows significant improvements in overall capabilities, from understanding Chinese and English text, executing complex instructions, handling subjective tasks, multilingual support to tool usage. The expansion is still ongoing and the official release will bring even more surprises,” Alibaba said in a press release.

Over the years, Qwen has helped Alibaba stand out in the global open-source AI ecosystem. These models have achieved over 20 million downloads and generated 100,000 derivative models on the Hugging Face platform. However, Qwen-3-Max-Preview is not yet open-source. Users can only access it via Alibaba Cloud or OpenRouter, similar to Qwen2.5-Max.

On social media, Binyuan Hui, an AI engineer at Alibaba, revealed that a "reasoning-capable" version of the model is still under development.

The cost of accessing Qwen-3-Max-Preview on Alibaba Cloud is also among the highest in the Qwen family. The starting price is $0.861 per million tokens input and $3.441 per million tokens output. Meanwhile, the Qwen3-235B-A22B-2507 version costs only $0.287 per million tokens input and $1.147 per million tokens output.

Source: https://znews.vn/alibaba-cong-bo-mo-hinh-thach-thuc-openai-post1583459.html