How Alibaba Creates Its Most Advanced AI Model Yet

Breakthrough in AI Efficiency: Alibaba's New Model Redefines the Future of Large Language Models

A recent breakthrough in artificial intelligence has captured the attention of experts worldwide, as companies seek to balance performance with cost efficiency. Alibaba Group Holding, a key player in China’s AI revolution, has introduced a new generation of foundation models that deliver strong performance while being significantly smaller and more cost-effective.

Alibaba Cloud, the company’s AI and cloud computing division, unveiled a new large language model on Friday that it claims marks "the future of efficient LLMs." This model, Qwen3-Next-80B-A3B, is nearly 13 times smaller than Alibaba’s largest AI model, which was released just a week prior. Despite its compact size, the model is considered one of Alibaba’s most advanced to date.

The innovation lies in its efficiency. According to developers, Qwen3-Next-80B-A3B performs 10 times faster in some tasks than the previous Qwen3-32B model, while reducing training costs by 90%. Emad Mostaque, co-founder of Stability AI, highlighted that the model outperformed many models from the previous year, despite an estimated training cost of less than US$500,000. For comparison, Google’s Gemini Ultra reportedly cost around US$191 million to train.

Artificial Analysis, a leading AI benchmarking firm, noted that Qwen3-Next-80B-A3B surpassed the latest versions of DeepSeek R1 and Moonshot AI’s Kimi-K2. The model’s success is attributed to a technique called "hybrid attention," which improves efficiency by making targeted adjustments to input data.

This approach addresses a common challenge in AI models: as input lengths increase, efficiency often declines. Traditional attention mechanisms involve trade-offs between accuracy and computational costs. However, Qwen3-Next-80B-A3B incorporates Gated DeltaNet, a technique proposed by researchers at MIT and Nvidia, to enhance attention without excessive costs.

Zhou Peilin, an AI researcher at the Hong Kong University of Science and Technology, explained that Gated DeltaNet allows the model to determine what information to retain and what to discard, resulting in a more accurate and cost-effective mechanism. Alibaba reported that the model achieved scores comparable to its most powerful model, Qwen3-235B-A22B-Thinking-2507, despite being smaller and more affordable.

Another key feature of the model is its use of the "mixture-of-experts" (MoE) architecture, which has driven efficiency gains in Chinese AI models over the past year. This architecture divides the model into specialized sub-networks that collaborate on tasks. Alibaba enhanced the sparsity of its MoE architecture, activating only 10 out of 512 experts at a time. This approach allowed the model to achieve parity with DeepSeek-V3.1, despite having just 3 billion active parameters compared to the latter’s 37 billion.

The efficiency gains are evident on Alibaba’s cloud platform, where the new model runs at a lower cost than the Qwen3-235B-2507 model, which contains 235 billion parameters. The trend toward smaller, more efficient models reflects growing concerns about the rising costs of scaling up large AI systems.

According to AI research firm Epoch AI, the most expensive training run to date was xAI’s Grok 4, costing US$490 million, with future runs expected to exceed US$1 billion by 2027. Meanwhile, researchers at Nvidia have advocated for small language models as the future of agentic AI due to their flexibility and efficiency.

Chinese AI giants are also pushing for broader adoption of their models by ensuring they can run on laptops and smartphones. Last month, Tencent Holdings launched four open-source AI models, each under 7 billion parameters, while Beijing-based start-up Z.ai released GLM 4.5 Air with 12 billion active parameters.

Alibaba’s Qwen3-Next-80B-A3B is now compact enough to operate on a single Nvidia H200 GPU, according to Artificial Analysis. On the open-source platform Hugging Face, the model quickly gained traction, amassing nearly 20,000 downloads within 24 hours of its launch.

Alibaba described the new architecture as a preview of its next generation of AI models. Tobias SchrOder, an AI researcher at Imperial College London, said the future of large language models will likely revolve around refining Alibaba’s approach to address training costs and efficiency, even if entirely different architectures emerge.