Alibaba's Qwen 3.8 open model with 2.4T parameters challenges frontier AI

Alibaba has announced the open release of its latest large language model, Qwen 3.8, featuring a massive 2.4 trillion parameters (2.4T). The Qwen team claims the model's performance is 'one of the highest currently available, comparable to frontier AI models and second only to Claude Fable 5.' While specific benchmark scores have not been published, a preview version called Qwen3.8-Max-Preview is already accessible via Qwen Studio and the Qwen Cloud token plan. Early users have demonstrated capabilities such as generating animated SVG images from natural language prompts (e.g., 'create an SVG of a pelican riding a bicycle'). The preview is actively improving daily, with the team reporting broad gains and significant upgrades in web frontend tasks.

This release signals Alibaba's return to an open model strategy after a brief detour with the closed Qwen 3.7. Previous open models like Qwen 3.5 and 3.6 spawned many community-driven optimizations, including the Bonsai 27B—a miniaturized version of Qwen 3.6 that runs locally on an iPhone using only 3.9GB of memory. Similar ecosystem activity is expected for Qwen 3.8. The move places Alibaba in the competitive open model landscape, where Chinese companies like Moonshot AI (with Kimi K3) have recently outperformed proprietary models from OpenAI and Anthropic in certain benchmarks. By keeping Qwen 3.8 open-weight, Alibaba invites widespread adoption and fine-tuning, potentially democratizing access to near-frontier AI capabilities.