Ethan Reid

Guest on “Deep Learning with Yacine”

An in-depth look at the principles behind our state-of-the-art segmentation model, and the architectural choices that got us there.

Speaker at Stanford's Frontier Research Club

Presentation and Q&A with Bay Area researchers, students, and investors.

SXSW: Seed Documentary Apple TV

HPE's 2017 documentary follows other founders and me through the startup process. Premiered at SXSW. Available on Apple TV.

Moondream Segmentation Model arXiv preprint

State-of-the-art referring segmentation model. A two-stage process: the model autoregressively generates a vector path, then refines the rasterization. Outperforms SAM3, Gemini 2.5 Pro, and Kimi SimpleSeg.

Lens

Distributed API for supervised and RL training of Moondream.

RefCOCO-M Benchmark

A cleaned version of the RefCOCO (UNC) validation split. Replaces the original instance masks with pixel-accurate masks and removes harmful samples, giving referring image segmentation a reliable benchmark to compete on.

Moondream 3 Model

Co-developed a 2A9B MoE VLM that exceeds GPT-5 and Claude 4 on perception and grounding tasks.

Moondream 2 Model 25M+ HF downloads

One of the first and most popular small VLMs. Over 30M downloads on Hugging Face.

Moondream 0.5B Model

The world's smallest VLM at the time of release. Trained through iterative pruning.