Alibaba has taken a new step in the global race for artificial intelligence with the launch of Qwen3, the third generation of its family of large-scale language models (LLM). In this way, the Chinese company not only introduces a more powerful technology but also strengthens its commitment to innovation in AI and continues to compete with the major industry giants today.
The new Qwen3 series includes eight different models, designed to adapt to all kinds of technological needs. Some are designed to operate quickly and lightly on devices like mobile phones or smart glasses, while others, much more powerful, are intended for complex tasks like data analysis, programming, or advanced automation.
Among them, two models called “Mixture-of-Experts” stand out, a technology that allows the use of only part of the model when needed, saving resources and reducing usage costs without losing quality in responses. This diversity aims to allow developers and companies to choose the model that best suits their needs, whether for mobile devices, smart glasses, autonomous vehicles, robots, or more traditional applications.
All these models have been released in open source and are available globally through platforms like Hugging Face, GitHub, and ModelScope.
One of the great innovations of Qwen3 is the introduction of hybrid reasoning, a system that allows the model to alternate between a “thinking mode” (for complex tasks like mathematics, programming, or logic) and a “non-thinking mode” (for simpler and faster responses). This ability to switch between modes aims to offer better computational efficiency without sacrificing quality. Thus, according to the company, some of these models achieve performance comparable to much larger dense models but with 10 times fewer active parameters.
The concept of “hybrid reasoning” is not completely new in AI literature, although very few companies have applied it as a practical application to a commercial model. Anthropic, for example, has Claude 3.7 Sonnet, which was the first AI model with hybrid reasoning. This model allows users to control the level of reasoning applied, offering everything from quick responses to detailed analysis according to the needs of the task.
For those accessing Qwen3 via API, Alibaba has included options for granular control of reasoning time, with a processing capacity of up to 38,000 tokens, allowing the balance between speed and depth to be adjusted.
Qwen3 has been trained with a gigantic database, with twice the tokens as its predecessor Qwen2.5. As a result, it offers substantial improvements in:
Additionally, its training included a four-phase process, from an initial “cold chain of thought” to reinforcement with reasoning and fusion of modes of thought.
For mas information about AI Agents: Welcome to the AI Agents Era: what they are and why they are changing everything
With over 300 million downloads and more than 100,000 derivative models created by the community, the Qwen series has become one of the most used in the world in the open-source AI environment. According to Alibaba, Qwen3 not only represents a technical advancement but also an inclusive and accessible proposal that aims to put the power of AI within reach of more developers. Soon, it will also be available via API from Alibaba’s Model Studio platform and already powers its flagship smart assistant, Quark.
You can try Qwen 3 directly from their website.
Image: Flux Schnell
Your email address will not be published. Required fields are marked *
Δ