© Alejandro Luengo/Pexels
After the start-up DeepSeek which shook the whole world and the stock market sector with its DeepSeek R1 model, it's the turn of Alibaba Cloud; a subsidiary of the Alibaba group; to offer its own chatbot. Called Qwen2.5-Max, it is part of the MoE (Mixture-of-Experts) models. These are neural network architectures that combine several “expert” sub-models. specialized to process different parts of the input data.
This allows it to claim a certain efficiency, particularly in natural language processing and has propelled it to the forefront of performance on several recognized benchmarks.
Qwen2.5-Max: the Chinese AI that rivals OpenAI and Anthropic
Trained on more than 20,000 billion tokens, Qwen2.5-Max then underwent a phase of “refining60~/em>”. A process that involves further training the model on a more specific and smaller dataset, in order to improve its performancefor a particular task. This combined supervised learning (SFT) and reinforcement learning based on human feedback (RLHF).
During the SFT phase, the model is given examples to learn the basics, and during the RLHF, it is “coached” by giving it human feedback to refine its skills.
Comparative evaluations place Qwen2.5-Max in a very favorable position compared to industry reference models(see below). On MMLU-Pro, which measures university-level knowledge, the model rivals GPT-4o (OpenAI) and Claude-3.5-Sonnet (Anthropic). More impressively, it outperforms DeepSeek V3 on several key metrics: Arena-Hard for alignment with human preferences, LiveBench for general abilities, and LiveCodeBench for programming skills.
Note that the latter is not open source, as is DeepSeek R1, and Alibaba Cloud did not measure its model against the latter.
Qwen2.5-Max outperforms its competitors on Arena-Hard and LiveBench. © Qwen
200% Deposit Bonus up to €3,000 180% First Deposit Bonus up to $20,000Two distinct versions of the model are available: a fundamental version and an “instruct” version ;» optimized for conversational applications and programming. It can be accessed either through Alibaba's cloud ecosystem, where it is referenced under the identifier “’ ;nbsp;qwen-max-2025-01-25 », via Qwen Chat (you will be asked to register for the’ use) or by its programming interface, rather intended for developers. It is compatible with the OpenAI API, a big plus for businesseswishing to integrate it into their architecture.
In use, what does it give for a typical user ? We tried it for a (short) half hour, here is what came out: Qwen2.5-Max is, in fact, quite efficient, answers questions without errors, and remains quite responsive. It has no problem performing complex calculations, has no problem generating/correcting computer code or reasoning on various subjects. Interestingly, answering tough questions about Chinese government policy is within his capabilities, unlike DeepSeek R1.
A rather neutral summary of the situation of the Uighurs in China. © Screenshot/Qwen Chat
The only major flaw that we were able to note is its capacity to generate images. Its competitors like DALL-E 3, Stable Diffusion or Midjourney still have a good head start. After several tests; on which the model has been churned for a relatively long time; the most beautiful image that we were able to take of it is the one located below. The margin of improvement is still immense, and the rendering is really average.
The result speaks for itself. © Screenshot/Qwen Chat
The goal of Alibaba Cloud, as she explains on her blog, is to “make our models smarter and able to reason like humans, or even better, through large-scale reinforcement learning. We want 8217;they explore areas of knowledge that are still unknown ». Qwen2.5-Max should therefore further progress in the coming months.
- Alibaba has launched Qwen2.5-Max, an advanced AI model based on MoE architecture, with optimal performance in language processing and programming.
- The model outperforms its main competitors on several benchmarks, but lags behind in image generation.
- It is free and accessible via the cloud, API or directly via the Qwen Chat site.
📍 To not miss any Presse-citron news, follow us on Google News and WhatsApp.
[ ]