Alibaba Unveils Latest AI Model Qwen 2.5-Max, Challenges DeepSeek Rising Influence

Chinese tech giant Alibaba has launched its latest AI model, Qwen 2.5-Max, in a challenge to the rising influence of DeepSeek.

The rollout which took place on the first day of the Lunar New Year, highlights the drive of Alibaba to stay competitive amid DeepSeek’s rapid ascent.

In an effort to reassert its dominance, Alibaba shared benchmark results suggesting that Queen 2.5-Max outperforms models from OpenAI and Anthropic in certain tasks. The company stated that its AI model achieves competitive performance against the top-tier models, referencing OpenAI’s GPT-4o, DeepSeek- V3, and Meta’s Llama-3.1, based on its compiled comparison using various benchmark tests.

Alibaba also announced that Qwen2.5-Max is now available in Qwen Chat, its chatbot-like web interface where users can interact directly with the model. Qwen 2.5-Max’s strong performance, according to Alibaba Cloud, shows that the expansion of data scale and model parameters can effectively improve intelligence of an AI model.

Qwen 2.5 consists of dense, decoder-only language models ranging from 0.5 billion to 72 billion parameters, available in base and instruct variants. The model is pre-trained on an 18-trillion-token dataset, supporting multilingual capabilities in 29 languages, including Chinese, English, Spanish, and Arabic. It can process up to 128,000 tokens in context and generate outputs of up to 8,000 tokens.

In addition to language capabilities, Qwen 2.5 improves document parsing, object detection, and video understanding. It enhances accuracy in identifying objects across multiple formats. Additionally, Qwen 2.5 comes with improved video understanding, taking a step up against the competition’s image understanding through ultra-long video processing and fine-grained video grounding.

Here are several other features that makes Qwen stand out:

Omni-modal Mastery

Can parse documents (tables, charts, handwriting), analyze long-form videos, and generate structured outputs (invoices, forms, etc.).

Training Efficiency

Pretrained on 18-20 trillion tokens with multimodal learning and fine-tuning for peak performance.

Benchmark Dominance

Outperforms GPT-4 and DeepSeek—delivering top-tier results with 40-60% fewer resources.

Notably, this model applies dynamic resolution and frame rate training for better temporal understanding, helping it extract event segments efficiently. It also integrates a streamlined vision encoder that improves both training and inference speeds.

Alibaba’s rollout of Qwen2.5, follows a market frenzy triggered by Chinese Al startup Deepseek, amid concerns over demand for the most advanced chips and data centers. Industry watchers believe that Alibaba’s announcement reflects the pressure that DeepSeek’s AI model is having not only on global rivals but domestic competition.

The release of DeepSeek’s AI assistant, powered by the DeepSeek-V3 model, as well also shocked Silicon Valley and caused tech shares to plunge, with the Chinese startup’s purportedly low development and usage costs prompting investors to question huge spending plans by leading AI firms in the United States.

Notably, DeepSeek-V3’s rise has influenced Al’s market trends, leading to upgrades from Alibaba, ByteDance, and Tencent. The startup open-source model is reported to have triggered an Al price war in China, prompting cost reductions among competitors. Alibaba’s latest model aims to provide an alternative with advanced multimodal functions and optimized inference speeds.

With China’s tech giants going heavy on AI, the rapid development of Al models in the country, signifies the increasing importance of the advanced technology. As the competition heats up the global tech ecosystem, we can expect further innovations and breakthroughs in the Al field.