Alibaba’s Qwen3.8-Max is now live for public access, and as the latest and most powerful model, it is touted to bring major upgrades in coding, research, long-horizon reasoning, and visual intelligence.
The Sparse Mixture-of-Experts (MoE) design powers the whole thing, and it only takes 95B parameters to activate during inference, and with a hybrid attention mechanism, this approach delivers frontier-level performance while keeping computational costs and latency significantly lower than traditional dense models of similar size.
The model currently ranks 5th on Text Arena and 2nd on Vision Arena as well as a 4th for Frontend Code Arena, with internal testing showing the AI model independently handled a real-world software engineering project over the course of 16 days, continuously generating code, testing, analyzing logs, and incorporating feedback to create an open-source self-evolving agent framework called oh-my-cli.
Legal document analysis and financial research to sports analytics, architectural 3D modeling, rehabilitation progress visualization, and even culinary concept development – all of these has seen some achievements as well, with the company claiming the model has demonstrated human-level capabilities in specialized research tasks, including outperforming human participants in the WWW2025 Multimodal Dialogue Intent Recognition Challenge.
Alibaba also introduced RecreationBench alongside the launch, a new benchmark designed to evaluate long-horizon application recreation. In testing, Qwen3.8-Max successfully reconstructed software applications from scratch in a black-box environment without internet access or source code, relying solely on visual feedback and iterative interaction. According to Alibaba, the results demonstrate the model’s growing capabilities as a hybrid AI agent capable of planning, adapting, and continuously improving over extended workflows.
Global developers can already access Qwen3.8-Max through Alibaba Cloud Model Studio APIs, while model weights are scheduled to be released next week. It’s also available to try through QwenWork.
