GitLab Transcend - Intelligent orchestration
Mini toolsMoondream 3 Preview
By Moondream
Moondream 3 Preview is Moondream’s newest vision-language model, designed around fast real-world visual intelligence rather than just generic image chat. It uses a fine-grained 9B mixture-of-experts architecture with 64 experts and 2B active parameters, supports both reasoning and non-reasoning modes, and expands context length from 2K to 32K. The release highlights strong visual reasoning, native pointing, improved object detection, better OCR, and more capable structured outputs, with an emphasis on trainability, low cost, and near-real-time inference for practical applications.
Multimodal
Gen 3
Overview
Moondream 3 Preview is a compact frontier-oriented vision-language model built for fast visual reasoning, grounding, OCR, object detection, pointing, and structured output. It uses a 9B MoE architecture with 2B active parameters and extends context length to 32K, aiming to deliver strong real-world vision performance while staying efficient and inexpensive to run.
About Moondream
State-of-the-art visual understanding at speeds that make continuous processing possible. Point, detect, count, and reason—without compromise.
Location:
Seattle, WA, US
Website:
moondream.ai
Tools using Moondream 3 Preview
No tools found for this model yet.
Last updated: April 2, 2026
