Molmo 2 8B

Molmo 2 8B is Ai2's flagship video-understanding model, released December 16, 2025. Built on Qwen3-8B with a SigLIP 2 vision encoder, it extends the original Molmo's image pointing to video pointing, multi-frame reasoning, object tracking, and counting, trained on about 9M videos with heavy human annotation.

Ai2 released the family in 8B, 4B, and 7B-O (OLMo-based, fully open stack) variants under Apache 2.0, publishing weights, training code, datasets, and evaluation tools. Vendor-reported results have the 8B beating Gemini 3 on video tracking while running on a single machine.