DeepSeek has been dethroned.

Alibaba's AI model, Qwen3, has passed R1 in LiveBench tests, which assess the capabilities of open-source models in programming, mathematics, and data analysis.

Qwen3, a new technology launched by Alibaba. Photo: SCMP .

According to the latest AI benchmark tests, Alibaba's newly released Qwen3 artificial intelligence model has surpassed DeepSeek's R1 to become the world's highest-rated open-source model.

Specifically, data from LiveBench, an independent platform that scores large language models (LLMs), the foundational technology for generative AI services like ChatGPT, shows that Qwen3 has surpassed R1 in the tests.

The assessment of the capabilities of open-source AI models includes programming, mathematics, data analysis, and language instruction.

Alibaba released its Qwen3 AI model series on April 28th. The company claims that this chatbot can rival, or even surpass, the best current models from OpenAI or Google in some cases.

With a size of up to 235 billion parameters, Qwen3 has performance comparable to DeepSeek-V2 and OpenAI GPT-4, which have approximately 236 billion and 175 billion parameters respectively. Users will soon be able to download it under open license on the AI ​​development platform Hugging Face and Github once this series of models is released.

The company stated that the Qwen3 collection includes hybrid models, meaning they can flexibly allocate time for reasoning to solve complex problems or quickly respond to simple requests. In this model, the reasoning capability allows for self-verification of information accuracy, but this comes with a significant time lag.

This design makes it easy for users to allocate budget appropriately to each specific task. Furthermore, this model also learns from many competitors around the world.

Using a "mixture of experts" (MoE) architecture similar to DeepSeek, Qwen3 can optimize computational performance while using only a fraction of the training cost. This method breaks down tasks into many separate parts and only requires a sufficient amount of data-intensive data to perform them.

According to the development team, Qwen3 supports up to 119 languages ​​and is trained on a dataset of nearly 36 trillion tokens, equivalent to 27 trillion words. Training data is sourced from various sources such as textbooks, question-answer sets, programming code, or even AI-generated data, etc.

Despite topping the open-source rankings, LiveBench's extended benchmark test shows that Qwen3 still lags behind the world's leading closed-source AI models. The most prominent among these are OpenAI's o3, Google's Gemini Pro 2.5, and Anthropic's Claude 3.7.

Currently, the most advanced OpenAI model supported by Microsoft, o3-mini high, tops the overall rankings of AI models worldwide.

Source: https://znews.vn/deepseek-bi-soan-ngoi-post1551500.html