On Codeforces, it's already making waves, surpassing OpenAI's o3-mini and Google's Gemini 2.5 Pro. Alibaba's new generation of artificial intelligence models, unveiled on Monday, April 27, is called Qwen3, and its version with the largest number of parameters (Qwen-3-235B-A22B) is reaching new heights. Looking in detail, we discover that it already performs better in terms of reasoning, on the BFCL test, the new benchmark for analyzing an AI's ability to reason about given problems.
Alongside DeepSeek, Alibaba is once again making its mark on the generative AI market, and several of its Qwen3 models are available to everyone, from platforms like Hugging Face and GitHub. The largest version is not yet available, but should be with open licenses. According to Alibaba, these models are "hybrid," meaning they can deliver on speed or prioritize the quality of reasoning. "We have seamlessly integrated thinking and non-thinking modes, giving users the necessary flexibility,” the Alibaba team responsible for Qwen explained in a blog post.
Alibaba Opens Up to Mixed Expert Architecture (MoE) for Its AI
In detail, Alibaba’s Qwen3 models are available in 119 languages, and have been trained on data at a scale of 36 trillion tokens. Previously, Qwen2 was unable to compete with available American AIs. With the largest model currently available to everyone, Qwen3 is now approaching R1 from the Chinese laboratory DeepSeek. Last January, Alibaba attempted to compete with OpenAI with Qwen2.5-Max, an advanced model comparable to GPT-4 or Anthropic's Claude-3.5-Sonnet.
To go further, Qwen3 has notably integrated a mixed expert architecture (MoE), a real breakthrough in neural models with a modular and specialized approach, which distributes a task into several sub-tasks, which will then be sent towards specialized models – the “expert” series – each designed to handle specific types of data or tasks.
Source: Tech Crunch
0 Comments