Alibaba's Qwen3.8-Max tops object detection with 2.4T-param MoE

Alibaba has quietly shipped Qwen3.8-Max, its most powerful vision-language model yet. Built on a Mixture-of-Experts architecture with 2.4 trillion total parameters (activating roughly 95 billion per query), the model accepts text, images, and video, and returns text. Roboflow's Piotr Skalski put it through a new VLM benchmark and found it outperformed every previous model on object detection — a first for a VLM in their testing. It handled satellite imagery, infrared shots, technical drawings, hand-drawn sketches, crowded scenes, and small objects without task-specific training.

What makes Qwen3.8-Max particularly interesting is its flexibility in prompting. Beyond traditional coordinate-based JSON output (best with XYXY normalized to 0-1000), the model supports box-based few-shot prompting. Marking a single pill with a bounding box and asking the model to find the rest yielded 80.8% mAP — slightly better than text-coordinate prompting. Positive and negative boxes also work to refine detection, though the model sometimes overfits to color or team affiliation based on the example box. In counting tasks, it tied Gemini 3.6 Flash, but still struggles in dense scenes, like miscounting metal brackets (10 vs. 12). The model is live via the Alibaba Cloud API and Roboflow Playground, with open weights and a smaller dense 27B version arriving August 12.