Alibaba unveils Qwen3.8-Max, its largest AI model yet, as China steps up race with global rivals

Alibaba on Monday introduced Qwen3.8-Max, calling it its "largest and most capable model to date" as the Chinese technology company looks to narrow the gap with the world's leading AI developers.

In a blog post, Alibaba said Qwen3.8-Max is built on a 2.4-trillion-parameter Mixture-of-Experts (MoE) architecture, with 95 billion parameters activated during inference. Instead of using the entire model for every request, only a small portion is activated, helping improve efficiency while reducing computing costs.

The company also said the model can handle a one-million-token context window, allowing it to analyse long documents, large code repositories and videos within a single prompt.

"We are very excited to release Qwen3.8-Max, our most capable foundation model to date," Alibaba said. It described the launch as "a significant leap forward" and said the model delivers stronger reasoning, coding and multimodal capabilities than previous versions.

Alibaba further claimed the model delivers "state-of-the-art performance across a broad range of benchmarks" while combining "high intelligence with practical deployment efficiency." Qwen3.8-Max will be available through Alibaba Cloud's Model Studio beginning next week.

How it compares with frontier models

Qwen3.8-Max is the biggest model Alibaba has built so far, but it is not the largest to have been announced publicly. That distinction still belongs to Moonshot AI's Kimi K3, which was launched last month with 2.8 trillion parameters.

Even so, model size no longer tells the whole story. AI companies are paying as much attention to how their models perform on reasoning, coding and multimodal tasks as they are to parameter counts. Cost and speed have become equally important, especially for enterprise customers.

Alibaba is making a similar case for Qwen3.8-Max. The company said it is the highest-ranked Chinese text model on Arena.AI and among the top multimodal models globally. Rather than focusing only on scale, it is highlighting benchmark performance, long-context capabilities and lower inference costs.

Comparing it with U.S. models is not straightforward. OpenAI, Anthropic and Google do not reveal the size of their latest models, including GPT-5.5, Claude Fable 5 and Gemini 3. Instead, those models are judged by how well they perform across reasoning, coding, tool use and real-world applications.

The launch also reflects how China's AI race is evolving; as companies competed to build larger models. Now the focus is shifting, with DeepSeek drawing attention for lowering inference costs, while Moonshot has pushed the limits of model scale. Alibaba, meanwhile, is trying to strike a balance between performance, efficiency and enterprise usability, a combination it believes will help Qwen3.8-Max compete with the latest frontier models.