Alibaba has introduced Qwen3.8-Max, the most advanced model in its Qwen family to date, expanding its capabilities in coding, research, multimodal reasoning, and long-horizon autonomous tasks.
Featuring 2.4 trillion parameters and supporting a context window of up to one million tokens, Qwen3.8-Max ranks fifth in Text Arena and second in Vision Arena, underscoring its competitive performance across language and vision benchmarks.
The model is now available to global developers through APIs on Alibaba Cloud Model Studio, while its model weights are scheduled for release next week. It can also be accessed via QwenWork, Alibaba’s workplace AI agent platform.
Efficient architecture for large-scale AI
Built on the foundation of Qwen 3.5, Qwen3.8-Max adopts a Sparse Mixture-of-Experts (MoE) architecture combined with a hybrid attention mechanism to improve efficiency.
Although the model contains 2.4 trillion parameters, it activates only 95 billion parameters during inference. This approach enables frontier-level performance while reducing computational costs and latency compared to traditional dense AI models of similar size.
Advanced coding and autonomous task execution
Qwen3.8-Max demonstrates strong capabilities in software development and long-term autonomous execution. Ranked fourth in Frontend Code Arena, the model can independently perform complex coding tasks over extended periods with minimal human intervention.
In internal testing, the AI autonomously completed a 16-day software engineering project, developing a self-evolving agent framework called “oh-my-cli.” Throughout the process, it continuously generated code, ran tests, analyzed logs, incorporated user feedback, and refined the project before it was fully open-sourced on GitHub.
Beyond software development, Alibaba said the model has shown the ability to improve upon published research methodologies and outperform human participants in the WWW2025 Multimodal Dialogue Intent Recognition Challenge, demonstrating advanced understanding of customer-service conversations and user intent.
Designed for real-world applications
Alibaba said Qwen3.8-Max is built to tackle a wide range of practical workloads across industries.
The model supports tasks including:
- Application and software design
- Legal document analysis
- Financial research
- Sports analytics
- Culinary concept development
- Rehabilitation progress visualization
- Architectural 3D modeling
The company added that the model combines reinforcement learning with increased computing power to strengthen autonomous planning, adaptive learning, and end-to-end task execution for complex, multi-step workflows.
Multimodal intelligence with visual understanding
As a multimodal foundation model, Qwen3.8-Max can process both text and visual content, transforming large amounts of information into searchable and interactive knowledge bases.
According to Alibaba, the model can analyze hundreds of pages of documents, entire television series, or 100-hour livestreams, converting them into structured information for easier retrieval and interaction.
Its visual capabilities also enable it to:
- Edit raw video footage into polished vlogs
- Generate educational animations from text prompts
- Rebuild complete frontend web applications from a single interface screenshot
- Convert 2D floor plans into detailed 3D interior designs
- Create interactive games from natural language instructions
To demonstrate these capabilities, Alibaba introduced RecreationBench, a benchmark designed to evaluate long-horizon application recreation. Working in a closed environment without internet access or source code, Qwen3.8-Max successfully reconstructed applications solely through interaction, visual feedback, and iterative refinement, highlighting its ability to solve complex tasks using multimodal reasoning.