Maestro is an open-source command-line toolkit from Roboflow that streamlines fine-tuning multimodal models, published on GitHub with about 2,700 stars. It wraps the training recipes for models such as PaliGemma 2, Florence-2 and Qwen2.5-VL into one consistent workflow.
Fine-tuning a different vision-language model usually means learning a new training script, data format and set of tricks. Maestro unifies that experience: one CLI and configuration style covers supervised fine-tuning for captioning, visual question answering and object detection across supported models, handling dataset preparation, training loops and evaluation. Cookbooks show end-to-end examples, including extracting structured JSON from images with Qwen2.5-VL. The project is Apache-2.0 licensed and actively maintained by Roboflow in 2026. The repository is part of Roboflow’s broader open-source ecosystem alongside tools such as supervision, and its recipes are maintained as new multimodal model generations are released. Because every model shares the same dataset conventions, teams can benchmark several architectures on identical data and pick the best fit before committing to a full training run.
Maestro is completely free and open source under the Apache-2.0 license. You run training on your own GPUs; Roboflow’s separate hosted platform is optional and not required.
Maestro is for computer vision engineers and ML teams who fine-tune multimodal models and want one repeatable workflow instead of per-model training scripts.