Alibaba Unveils Smaller AI Model Matching Larger Ones Efficiently

Photo:Reuters

Alibaba, one of the biggest technology companies in China, has introduced a new AI model that is smaller, faster, and more cost-efficient but still performs at a level similar to their largest AI models. This is a significant step forward in artificial intelligence because it shows that you don’t always need huge models to achieve powerful results.

The new AI model, called Qwen3-Next-80B-A3B, was launched by Alibaba’s cloud computing and AI division. It is nearly 13 times smaller than Alibaba’s largest AI model released just a week before. Despite its smaller size, developers say it is one of Alibaba’s best models yet.

The model is much faster and cheaper to run. For certain tasks, it works 10 times faster than the previous model, Qwen3-32B, released in April. Training this AI model cost 90% less than larger models, making it much more efficient. Experts estimate the total training cost was less than US$500,000, which is extremely low compared to models like Google’s Gemini Ultra that cost US$191 million to train.

One reason for this success is a new technology called “hybrid attention” combined with “Gated DeltaNet”, a technique first suggested by researchers from MIT and Nvidia. Normally, AI models spend a lot of computing power deciding which parts of the input information are important. This is called the “attention” mechanism. It becomes very costly when models have to process long inputs.

The Gated DeltaNet method improves attention in a smarter way. It helps the AI figure out which information to keep and which to ignore, making the model both accurate and cheaper to train. According to testing, Qwen3-Next-80B-A3B performs almost as well as Alibaba’s most powerful model, Qwen3-235B-A22B, even though it is much smaller.

Stay informed with the latest news. Follow DXB News Network on WhatsApp Channel 

In addition, the model uses a special architecture called “mixture-of-experts” (MoE). This splits the AI into smaller networks called experts that specialize in different parts of the data. Alibaba’s model has 512 experts, but only 10 are active at a time, which makes it highly efficient. For comparison, other AI models have fewer experts and use more resources, but this new setup helps the model work like larger models while using fewer resources.

This new approach shows a shift in AI development, where companies focus on smaller, more efficient models instead of simply making bigger ones. This trend is growing worldwide, as large AI models are becoming very expensive to train. Some AI models now cost hundreds of millions of dollars to train, and future projects could surpass US$1 billion.

Alibaba’s new model is compact enough to run on a single high-end graphics card, making it possible for developers to use it on personal computers and laptops. This is part of a broader push by Chinese AI companies to make AI more accessible and practical for everyday use.

After its release, the model quickly became popular on developer platforms, with almost 20,000 downloads in 24 hours. Experts believe that this model is not just a one-time improvement, but a preview of how the next generation of AI models will work: smaller, faster, and more cost-efficient without losing performance.

Overall, Alibaba’s launch marks a new era in AI technology, proving that innovation can make artificial intelligence powerful, affordable, and accessible. This development could help more people and businesses use advanced AI tools without the huge costs usually associated with large AI models.