Model
Inference & Serving
Molmo
1Fully 2open 3multimodal 4models 5from 6Ai2
the six Ws · specification
W1
Who
The Allen Institute for AI (Ai2) built the Molmo family of multimodal models.
W2
What
Open vision-language models trained entirely on open datasets, including the PixMo image-caption corpus.
W3
Where
Published on Hugging Face under the allenai organization.
W4
When
Molmo was released in September 2024.
W5
Why
It aimed to make state-of-the-art multimodal AI fully open, including data, unlike most closed VLM competitors.
W6
With
Runs on PyTorch and transformers, built on Qwen2 and OLMo language model backbones with a CLIP-based vision encoder.
W7
Watch
Fully open release (weights, code, and training data) by Ai2 under Apache 2.0 license.