Molmo 2 4B

Molmo 2 4B is the efficiency-optimized member of the Molmo 2 family, released by the Allen Institute for AI (Ai2) in December 2025. It extends the original Molmo image-understanding model into video: the model tracks objects over time, pinpoints where and when events occur in a video, and reasons across multiple images at once. It is built on a Qwen 3 4B language backbone with a SigLIP 2 vision encoder.

Ai2 released the full Molmo 2 family (4B, 8B, and the fully open Molmo 2-O 7B) with open weights, open training data, and open recipes under Apache 2.0. The release notes report that the 8B sibling matches or beats much larger proprietary models on video tracking benchmarks while training on far less data; the 4B variant trades some accuracy for cheaper, faster deployment. Exact release dating varies across sources (December 2025 versus January 2026 listings), so the date below is era-level.