moondream1: a 1.6B VLM that runs on CPU
In one sentence On January 20, 2024 independent developer vikhyatk releases moondream1, a 1.6-billion-parameter vision-language model built with SigLIP, Phi-1.5 and the LLaVA dataset. Research-only license, commercial use prohibited.
Models that look at an image and answer questions about it are usually large and need powerful GPUs. moondream1 tries the opposite route: a small model, about 1.6 billion parameters, which the sample script runs on CPU.
You give it an image and a question ("what color is the train?", "what is the book's title?") and it answers in English. In the author's published benchmarks it stays below much larger models such as LLaVA-1.5 7B and 13B, but beats other small models on some tests.
The model is released for research use only: commercial use is not allowed.
Companies
vikhyatk (independent)
Tools
moondream1
Tags
Sources