Chinese company Alibaba Cloud has unveiled its most advanced language model of artificial intelligence Qwen 2.5-Max. According to the developers, it is superior to the model of DeepSeek.
What Is the Innovation
The Alibaba model cannot be called a “reasoning model”, unlike DeepSeek-R1, which became a sensation. That is, Qwen 2.5-Max does not use a “chain of thought” to form more meaningful and justified answers. Therefore, it is more correct to compare it with the DeepSeek-V3, Llama-3.1 and other models, which are the basis of “reasoning models”.
The Qwen 2.5-Max benchmark results, compared to DeepSeek-V3 and other models, show the following:
- Arena-Hard (user preference): Qwen 2.5-Max scored 89.4, ahead of DeepSeek-V3 (85.5) and Claude 3.5 Sonnet (85.2).
- MMLU-Pro (university-level problems): Qwen 2.5-Max scored 76.1, slightly ahead of DeepSeek V3 (75.9), but behind the leader Claude 3.5 Sonnet (78.0) and the second-placed GPT-4o (77.0).
- LiveCodeBench (programming): Qwen 2.5-Max scored 38.7, which is roughly in line with DeepSeek-V3 (37.6), but slightly behind Claude 3.5 Sonnet (38.9).
- LiveBench (overall real-world AI performance): Qwen 2.5-Max took first place with 62.2, beating DeepSeek-V3 (60.5) and Claude 3.5 Sonnet (60.3).
The testing confirms that the Qwen 2.5-Max is capable of competing with and potentially outperforming leading models from other companies.
What the Model Was Trained On
Like other companies, Alibaba does not disclose the data sources used to train Qwen 2.5-Max. The model was pre-trained on over 20 trillion tokens — about 15 trillion words! Qwen 2.5-Max uses a Mixture-of-Experts (MoE) architecture, the same training technology used by DeepSeek V3. This approach allows the model to scale while keeping the computational cost reasonable.
The sampling mechanism makes processing large amounts of data more efficient without incurring excessive computational costs. After pre-training, the model undergoes further tuning using Supervised Fine-Tuning (SFT) and Reinforcement Learning from Human Feedback (RLHF). This means that humans test and adjust it, helping it become more accurate.
How to Use the Model
Qwen 2.5-Max is not considered an open source model and is available for use through the Qwen Chat platform and API on Alibaba Cloud. Users can interact with it directly through the web interface. They can integrate its capabilities into their applications through the API. So far, the DeepSeek-V3 is the cheapest of the three. It makes it attractive, especially for tasks that require processing large amounts of text. But since the Qwen 2.5-Max scores well in benchmarks, it may be competitive in tasks where accuracy is more important than price.
Hell Spin Casino Live Gaming is another example of how AI advancements influence different industries, including online entertainment. The integration of AI-driven models helps platforms like HellSpin casino offer a more interactive and personalized experience for players.
What Does the New Model Give to the Market and Ordinary Users
For the AI industry, this means that competition between companies developing AI will increase. But also that the market volume will potentially grow. More and more clients will be able to afford the implementation of AI.
As training becomes cheaper, the number of startups capable of developing more specialized models for specific tasks may increase.
Increased specialization and diversity will lead to a decrease in the share of each individual large company, for example, for OpenAI. This trend was already observed in 2024, only from large cloud providers expanding their offering and market share. Reducing barriers to entry (the cost of resources for training) will only accelerate the trend, but from new specialized players. As for business and ordinary people, AI tools will become more accessible. End users will be able to “juggle” models, choosing the best option for solving a particular problem.
Will China Become the First in the AI Race
The launches of DeepSeek and Qwen 2.5-Max demonstrate China’s progress in the field of artificial intelligence. It shows the ability of Chinese companies to develop models comparable to and even superior to their US counterparts. But it is too early to talk about a winner in this race, as competitiveness will be affected by many factors. Currently, the US dominates the production of more advanced chips. China produces many times more electricity, which is necessary for scaling AI data centers. Both countries understand that AI and robotics will determine the competitiveness of the economy in the near future. Thus, they are investing in infrastructure and development of such technologies.
However it is not true that only Chinese companies can achieve breakthroughs in reducing the cost of training AI models. Just a week before the release of the DeepSeek model, a team from the Sky Computing lab at UC Berkeley presented a “reasoning model” that cost less than $450 to train. Despite the low cost, it showed impressive results. So the emergence of barriers such as high cost or lack of AI chips only encourages researchers and startups to look for ways to overcome them. In the future, we may see other breakthroughs that will completely overturn the rules that seemed immutable just yesterday.