ALIBABA recently launched Qwen3.8-Max, the most powerful model in its Qwen series to date. Boasting 2.4 trillion parameters and supporting a context window of up to 1 million tokens, Qwen3.8-Max ranks fifth in Text Arena and second in Vision Arena, while exhibiting advanced capabilities in coding, real-life work, research, and long-horizon tasks. Operating as a multimodal foundation, Qwen3.8-Max also supports visual intelligence.
The model is now accessible via APIs on the Alibaba Cloud Model Studio for global developers, with model weights scheduled for release next week. It can also be experienced on QwenWork, Alibaba’s latest all-in-one workplace AI agent platform.
Qwen3.8-Max ranks fifth in Text Arena and second in Vision Arena
Innovative Architecture: Balancing Massive Scale with High Efficiency
Built upon the robust foundation of Qwen 3.5, Qwen3.8-Max features a cutting-edge Sparse Mixture-of-Experts (MoE) architecture combined with a hybrid attention mechanism. This design achieves a critical balance between massive scale and inference efficiency. Despite its total size of 2.4 trillion parameters, Qwen3.8-Max activates just 95 billion parameters. This innovative framework allows the model to deliver frontier-level intelligence for complex tasks while significantly reducing computational costs and latency compared to traditional dense models of similar scale.
Frontier-Level Capabilities in Autonomous Coding, Real-world Applications, and Long-Horizon Tasks
Ranked fourth in Frontend Code Arena, Qwen3.8-Max demonstrates exceptional proficiency in autonomous coding and long-horizon execution, operating independently over extended periods without human intervention. In internal testing, the model autonomously executed a real-world software engineering project over a 16-day period. Tasked with creating a self-evolving agent framework from scratch, Qwen3.8-Max established an engineering loop that synthesized user feedback, community best practices, and self-test data. By continuously iterating through code generation, testing, previewing, and log analysis, it produced “oh-my-cli“, a self-evolving agent framework that has been fully open-sourced on GitHub.
Beyond coding automation, Qwen3.8-Max drives genuine innovation in highly specialized fields. For example, the model not only reproduced published research experiments but also engineered novel methodologies that outperformed the original papers. Proving its human-level competence, Qwen3.8-Max outperformed human participants in the WWW2025 Multimodal Dialogue Intent Recognition Challenge, showcasing an unmatched ability to analyze customer-service transcripts and accurately interpret complex user demands.
Additionally, Qwen3.8-Max is engineered to manage intricate, real-world workloads, spanning application design, legal document review, sports analytics, financial research, culinary concept development, rehabilitation progress visualization, and architectural 3D modeling. By jointly scaling reinforcement learning environments and compute, the model significantly enhances general operational competence across mainstream agent frameworks. When addressing highly complex, multi-constraint, and long-horizon challenges, Qwen3.8-Max also exhibits elite system-level autonomous planning and end-to-end, closed-loop adaptive learning capabilities.
Visual Intelligence and Multimodal Agent Capabilities
Operating as a multimodal foundation, Qwen3.8-Max supports visual intelligence, transforming static inputs into dynamic, interactive knowledge structures. The model can seamlessly ingest hundred-page documents, full television series, or 100-hour livestreams, converting them into searchable, interactive knowledge bases.
Qwen3.8-Max excels at continuous execution and creation based on real-time visual feedback. Its versatile capabilities include editing raw personal footage into professional vlogs, generating immersive educational animations from text prompts, reconstructing complete frontend web projects from a single user-interface screenshot, transforming 2D floor plans into detailed 3D interior visualizations, and building interactive games directly from natural language requests.
To demonstrate this, the team introduced RecreationBench, a long-horizon application-recreation benchmark. Operating entirely in a black-box environment with no internet access or source code visibility, Qwen3.8-Max autonomously reconstructed the applications from scratch by evaluating live applications purely through interaction and feedback. It shows that by relying entirely on interactive evaluation and visual feedback, the model demonstrated frontier hybrid agent capabilities, advancing visual coding through iterative development.
