Alibaba has introduced Qwen3.8-Max — the most powerful model in the Qwen lineup

Alibaba announced the launch of Qwen3.8-Max — the largest and most capable model in the Qwen series. The model features 2.4 trillion parameters and supports a context window of up to one million tokens. Qwen3.8-Max ranked fifth in the Text Arena leaderboard and second in the Vision Arena leaderboard, while also demonstrating strong performance in coding tasks, real-world scenarios, research, and long-horizon autonomous task execution.

The model is already available via API on Alibaba Cloud Model Studio for developers worldwide, with model weights to be released next week. Qwen3.8-Max can also be tested on QwenWork — Alibaba’s new platform for AI agent workflows.

Qwen3.8-Max is built upon Qwen 3.5 and utilizes a sparse mixture-of-experts architecture combined with a hybrid attention mechanism. With a total size of 2.4 trillion parameters, the model activates only 95 billion parameters per token, enabling high performance while reducing computational costs and latency compared to dense models of similar scale.

Qwen3.8-Max ranked fourth in the Frontend Code Arena leaderboard and achieved strong results in autonomous coding and long-horizon task execution. In internal testing, the model independently managed a real engineering project for 16 days, building from scratch an self-evolving agent framework called oh-my-cli, which was subsequently released on GitHub.

The model also outperformed human participants in the WWW2025 Multimodal Dialogue Intent Recognition Challenge, demonstrating the ability to analyze customer support transcripts and accurately interpret complex user requests.

Among its capabilities are converting amateur videos into professional vlogs, generating educational animations from text prompts, reconstructing full frontend projects from a single UI screenshot, transforming 2D floor plans into 3D interior visualizations, and creating games from text descriptions.

To evaluate these capabilities, the development team introduced the RecreationBench benchmark, under which the model reconstructed applications from scratch relying solely on interaction with them and visual feedback, without access to the internet or source code.

Source: alizila.com