Molmo2

Molmo2 is the Allen Institute for AI's second-generation open vision-language family, released December 16, 2025 and documented in a CVPR 2026 paper (arXiv 2601.10611). Three variants cover the spread: Molmo2-4B and Molmo2-8B on Qwen3 backbones, and Molmo2-O-7B on Ai2's own OLMo for a fully open end-to-end stack. Beyond image and multi-image understanding, the headline is video: spatio-temporal grounding that pinpoints where and when objects and events occur — points, object tracks, and grounded chain-of-thought.

This is the rare 'actually open' release: weights, training code, and all training data are public under Apache 2.0, with the data deliberately collected without distilling closed VLMs. Ai2 reports state-of-the-art results among open models on video grounding and captioning, with human-preference evaluations on par with proprietary systems (provider-reported). Weak spots per third-party notes: long (10min+) video, dense document OCR, and multimodal math reasoning.

Provider-reported: SOTA among open models on video grounding/captioning; human-preference evals on par with proprietary systems.

Independent: GeekWire and press coverage confirm the release scope; no independent benchmark reproduction found.