Alibaba Launches Qwen3.8-Max AI Model

Alibaba has officially launched Qwen3.8-Max, the latest and most advanced model in its Qwen artificial intelligence series, introducing a multimodal foundation model designed for complex coding, research, workplace and long-horizon tasks.

The new model features 2.4 trillion parameters and supports a context window of up to 1 million tokens. According to Alibaba, Qwen3.8-Max ranks fifth in Text Arena and second in Vision Arena, while also delivering enhanced capabilities in autonomous coding, real-world applications and visual intelligence.

Qwen3.8-Max is now available through APIs via Alibaba Cloud Model Studio for developers worldwide, with model weights scheduled for release next week. The model is also available through QwenWork, Alibaba's latest all-in-one workplace AI agent platform.

Built on the foundation of Qwen 3.5, Qwen3.8-Max uses a Sparse Mixture-of-Experts (MoE) architecture combined with a hybrid attention mechanism to balance large-scale intelligence with improved efficiency.

Although the model contains 2.4 trillion total parameters, Alibaba says it activates only 95 billion parameters during operation. The company claims this approach enables frontier-level performance on complex tasks while reducing computational requirements and latency compared with traditional dense models of a similar scale.

Qwen3.8-Max has been ranked fourth in Fronted Code Arena, highlighting its performance in autonomous coding and extended task execution, according to Alibaba.

In internal testing, the model independently completed a real-world software engineering project over a 16-day period. The task involved creating a self-evolving agent framework from the ground up, with Qwen3.8-Max developing an engineering cycle that incorporated user feedback, community practices and self-testing data.

Through repeated cycles of code generation, testing, previewing and analysis, the model created "oh-my-cli", a self-evolving agent framework that has been released as an open-source project on GitHub.

Alibaba said the model has also demonstrated capabilities beyond software development, including reproducing published research experiments, developing new methodologies and supporting specialist applications. In one example, Qwen3.8-Max outperformed human participants in the WWW2025 Multimodal Dialogue Intent Recognition Challenge, analysing customer service conversations and interpreting complex user requirements.

The model has also been designed for a wide range of real-world applications, including application development, legal document analysis, sports analytics, financial research, culinary design, rehabilitation visualisation and architectural 3D modelling.

As a multimodal foundation model, Qwen3.8-Max supports visual intelligence, enabling users to transform static content into interactive knowledge structures.

Alibaba said the model can process large-scale visual and document-based information, including hundreds of pages of documents, complete television series and 100-hour livestreams, converting them into searchable and interactive knowledge resources.

The model can also use visual feedback for continuous creation and execution, with applications including turning personal footage into professional videos, generating educational animations from text prompts, recreating frontend web projects from interface screenshots, converting 2D floor plans into detailed 3D interior visualisations and creating interactive games from natural language instructions.

To demonstrate these capabilities, Alibaba introduced RecreationBench, a long-horizon application recreation benchmark. In a closed environment without internet access or source code visibility, Qwen3.8-Max recreated applications by interacting with live versions and using visual feedback to guide development.