Alibaba Unveils Qwen 2.5-Max, the New AI Model That Challenges DeepSeek and OpenAI | Large language models courses | Best llm training dataset github | Top 20 most popular large language models in the world | Turtles AI

Alibaba Unveils Qwen 2.5-Max, the New AI Model That Challenges DeepSeek and OpenAI
Alibaba Challenges DeepSeek: New Qwen 2.5-Max Model Aims to Redefine AI Standards
Isabella V29 January 2025

 

Alibaba has released Qwen 2.5-Max, an advanced AI model that the company says outperforms DeepSeek-V3, GPT-4o, and Llama-3.1-405B in multiple benchmarks. The model uses a Mixture-of-Experts (MoE) architecture and has been trained on over 20 trillion tokens, with advanced refinement techniques. The launch, on the first day of the Lunar New Year, highlights the growing competition in China’s AI industry, accelerated by DeepSeek’s rapid progress.

Key Points:

  • Advanced Performance: Qwen 2.5-Max outperforms DeepSeek-V3 and other benchmarks on critical tests.
  • Innovative Training: The model uses a huge amount of data and advanced methodologies such as SFT and RLHF.
  • Direct Challenge to AI Giants: Alibaba Enters the Competition with OpenAI, Meta, and DeepSeek.
  • API Availability: Developers can access Qwen 2.5-Max via Alibaba Cloud.

On Wednesday, Alibaba announced the release of Qwen 2.5-Max, a new version of its AI model, aiming to redefine the industry standard. Strategically launched on the first day of the Lunar New Year, the model is a direct response to the rise of DeepSeek, a Chinese startup that disrupted the market with the release of its DeepSeek-V3 and subsequent DeepSeek-R1. DeepSeek’s rapid growth has sparked an innovation race among major players in the industry, prompting giants like Alibaba and ByteDance to accelerate the development of their own models.

Qwen 2.5-Max stands out for its use of a Mixture-of-Experts (MoE) architecture, a technique that optimizes the use of computational resources, improving performance without exponentially increasing processing costs. The model was trained on over 20 trillion tokens and refined through Supervised Fine-Tuning (SFT) and Reinforcement Learning from Human Feedback (RLHF), two methodologies that sharpen the AI’s ability to understand and respond.

Comparative evaluations conducted by Alibaba place Qwen 2.5-Max at the top of several leading benchmarks. The model outperformed DeepSeek-V3, GPT-4o, and Claude-3.5-Sonnet on tests such as Arena-Hard, LiveBench, LiveCodeBench, and GPQA-Diamond. It also showed competitive performance in other areas, including MMLU-Pro, a test that measures university-level academic knowledge. When compared to available base models, Qwen 2.5-Max outperformed DeepSeek-V3 and Llama-3.1-405B, confirming its position among the best open-weight solutions.

Alibaba has made Qwen 2.5-Max available via Qwen Chat, giving users the ability to interact directly with the model. For developers, the model is accessible through the API on Alibaba Cloud, after registering and activating the Model Studio service. The company emphasizes that the continued increase in computational capacity and the refinement of post-training techniques will lead to further improvements in future iterations of the model.

Alibaba’s ambition with Qwen 2.5-Max is to push the limits of current AI, improving the reasoning and problem-solving capabilities of language models.

The use of reinforcement learning at scale aims to create systems that can surpass human intelligence in certain tasks, opening the way to new possibilities for discovery and innovation.