Molmo by Ai2 at a glance

Molmo is a family of vision-language models from the Allen Institute for AI (Ai2). It combines an image encoder with a language model so users can ask natural-language questions about visual inputs.

Its distinctive capability is visual grounding: rather than only describing an object, Molmo can return points that show what part of the image supports the answer. That is useful for counting, spatial reference and debugging what the model appears to attend to.

The original family includes MolmoE-1B, Molmo-7B-O, Molmo-7B-D and Molmo-72B. The 7B-D model powers the preserved public demo, while 7B-O uses Ai2's OLMo backbone and is the more open 7B option.

Molmo 2 is now the current general-purpose successor. It extends the original image workflow to multiple images and videos, with question answering, dense captioning, pointing, timestamps and persistent object tracking.

Molmo 2 comes in 4B, 8B and 7B-O variants. Ai2 describes 4B as the efficient option, 8B as its best overall choice for short video and image tasks, and 7B-O as the model with an OLMo backbone and a fully open end-to-end model flow.

For video, Molmo 2 can sample frames, identify when and where events occur, count by pointing and track objects across occlusion. Ai2's own evaluation also shows meaningful headroom: its reported video-grounded counting accuracy remained below 40%.

Developers can run the model with Hugging Face Transformers, use vLLM support through a nightly build, fine-tune the released code or evaluate it in the Ai2 Playground. These routes require substantially different engineering and privacy assumptions.

The model weights and core repositories use Apache 2.0, but licensing is layered. Ai2 says Molmo 2 is intended for research and education and warns that some third-party datasets in the training mix are limited to academic or non-commercial research.

A model that points to an object is still not proving the answer. Grounding coordinates can be misplaced, counts can be wrong, text in images can be misread and video events can be missed between sampled frames.

Running locally can keep inputs inside the operator's infrastructure, but it also transfers responsibility for access controls, logs, updates, model safeguards, GPU capacity and incident response to that operator.

Molmo should be treated as an inspectable component rather than a turnkey decision system. Human review and task-specific evaluation are required before using its output for accessibility, moderation, medical images, robotics, surveillance, hiring or industrial quality control.