Molmo 2 doesn't just answer questions about clips—it searches & points. The model returns coordinates & timestamps over videos + images, powering QA, counting, dense captioning, artifact detection, & subtitle-aware analysis. You can see exactly how it reasoned.

Last year Molmo set SOTA on image benchmarks + pioneered image pointing. Millions of downloads later, Molmo 2 brings Molmo’s grounded multimodal capabilities to video 🎥—and leads many open models on challenging industry video benchmarks. 🧵

We’re releasing three Molmo 2 variants that run at 2 FPS/8 FPS, depending on your performance needs: 🔹 Molmo 2 (8B): Qwen 3 backbone, best overall grounding + QA 🔹 Molmo 2 (4B): Qwen 3 backbone, fast + efficient 🔹 Molmo 2-O (7B): Olmo backbone, fully open model flow

🎯 Demo: counting objects & actions Show Molmo 2 a clip and ask “How many times does the ball hit the ground?” It returns a count plus space–time pointers (pixel coordinates + timestamps) for each event, so you can verify every answer.

❓ Demo: ask-it-anything video QA Show Molmo 2 a long clip and ask a complex question – “Why does the player change strategy here?” It reasons over the whole timeline & can point to the exact moments that support its answer.

📍 Demo: object tracking Ask Molmo 2 to follow an object—“Follow the red race car.” It places a pink dot on the object and tracks it across frames, returning coordinates over time that you can plug into your workflow.

For Molmo 2, we've significantly upgraded the Ai2 Playground. 🛠️ You can now upload a video or multiple images to try out video summarization, tracking, & counting—giving you a clear view of exactly where the model is looking.

Molmo 2: open, customizable video + image understanding. ▶️ Try in the Playground: ⬇️ Download: 📝 Blog: 📑 Report: 💻 API coming soon

def taking this for a spin

Providing the spatial-temporal index of the output is not the same as illuminating the opaque inductive logic that generated it.