Moondream: how does a tiny vision model slap so hard?
AI Engineer World's Fair 2024 · 19:26
Computer vision and vision-language models
M87 Labs builds Moondream, a family of vision-language models for developers creating applications that interpret images. The models answer visual questions, generate captions, detect objects and segment image regions. Lens provides fine-tuning through an API, while Photon runs inference locally and Moondream Cloud offers hosted inference. Developers can adapt models to their own data and deploy across cloud servers, workstations and edge devices.
Founded in 2024 by Jay and Vik Korrapati after eight years working together at AWS, the San Francisco company focuses on computer vision; Korrapati serves as co-founder and CTO. Its technical work includes Moondream Segmentation, which turns descriptions into object masks by generating vector paths and refining their rasterized boundaries. Reinforcement learning directly optimizes mask quality. The associated research also introduced RefCOCO-M, a cleaned validation dataset with more accurate object boundaries.
As of August 2026, the company reported more than 5 million monthly downloads of Moondream. Its business combines free models and Photon software with token-based charges for Cloud inference and Lens training. Paid plans add collaboration and support, including shared fine-tunes and custom deployment assistance. Hosting the models as a cloud service requires a license, an important condition for developers considering that distribution model.
AI Engineer World's Fair 2024 · 19:26
Affiliations reflect their AIE appearances, not necessarily current employment.
Korrapati describes how poorly matched GPT-4-generated synthetic training data can increase hallucinations, highlighting the importance of representative training distributions.
Affiliations reflect each recorded session, not necessarily current employment.