Molmo 72B on real-time camera streams
Molmo 72B · Allen Institute for AI · VLM · Open-weight
Molmo 72B is the Allen Institute for AI's open-weight vision-language model, trained on the fully open PixMo dataset and built to lead on pointing and visual grounding rather than just general description. Instead of only describing an image, it can indicate exactly where an object is within a frame, which is valuable for applications that need to act on what a camera sees. Molmo 72B is not currently in Overshoot's live model catalog. Models that are live in the catalog answer questions about a live WebRTC stream through an OpenAI-compatible chat-completions request.
Molmo 72B is not currently in the live Overshoot model catalog (verified against the live model catalog on 2026-07-14). Availability changes over time - query GET /v1beta/models for the current list.
- Developer
- Allen Institute for AI
- Parameters
- 72B
- Context window
- 4K tokens
- License
- Apache 2.0
- Released
- Sep 2024
- Inputs
- Text, images, video frames
- Overshoot availability
- Not in live catalogas of 2026-07-14
What Molmo 72B is good at
Molmo 72B's standout capability is pointing: given a question about a frame, it can return the specific location of the object or region in question rather than only a text description. That grounding makes it a strong fit for UI automation, where an agent needs to know exactly where to click, and for robotics, where a system needs a precise target location in the frame.
It was trained entirely on the openly released PixMo dataset rather than distilled from a closed model, which the Allen Institute for AI treats as important for reproducibility and for understanding exactly what the model has and has not seen.
- Precise pointing and grounding at specific frame locations
- UI automation that needs an exact click or interaction target
- Robotics perception that needs a grounded location, not just a description
Molmo 72B and Overshoot's streaming workflow
Molmo 72B is not currently in Overshoot's live model catalog. Models that are live read footage the same way: publish a camera or screen share over WebRTC to a Stream, and Overshoot retains 600 seconds of frame history behind it. A chat-completions request references that footage with an ovs:// URL, anchored to the latest frame, an exact timestamp, or a recent segment.
If Molmo 72B were served through the API, its pointing and grounding answers would stream back token by token over SSE. Its 4K-token context window is narrow compared with newer models, so applications would typically keep each request focused on the current frame or a short recent segment rather than a long running history.
Molmo 72B versus general-purpose VLMs
Most vision-language models describe an image well but are imprecise about exact locations within it. Molmo 72B is built specifically to close that gap, trading some general-purpose context length and conversational range for class-leading pointing accuracy. Teams building UI agents or robotics perception on live video reach for Molmo 72B where the task is fundamentally about grounding a location, and use a broader model like Qwen2.5-VL 72B when the task is general visual question answering instead.
Frequently asked questions
Can Molmo 72B analyze live video?
Molmo 72B works frame by frame, including pointing to specific locations within an image, which suits sampled live video frames. It is not currently in Overshoot's live model catalog, so it cannot be called against Overshoot streams today. Catalog models reference the latest frame or a recent segment with an ovs:// URL and stream answers back over SSE.
Is Molmo 72B open source?
Molmo 72B is released under the Apache 2.0 license with fully open weights, and it was trained on the openly released PixMo dataset rather than a closed one.
Is Molmo 72B available on Overshoot?
Not currently. Molmo 72B is not in Overshoot's live model catalog. The catalog changes over time, so check GET /v1beta/models for the current list of models available through the API.
What is Molmo 72B best used for?
Molmo 72B is best suited to tasks that need precise pointing or grounding, such as UI automation that must identify an exact click target or robotics perception that needs a specific location in the frame. For general visual question answering without a grounding requirement, a broader model is often a better fit.