# Ai2 released Molmo 2 with open video-grounding models, data and checkpoints Ai2 released Molmo 2 on December 11, 2025, adding 4B, 8B and fully open 7B video-language variants with pointing and tracking artifacts. Canonical: https://rangoon.ai/insights/archive-ai2-molmo2-video-grounding-december-2025/ Author: Rangoon Editorial (https://rangoon.ai/insights/editorial/) Published: 2026-10-24 Updated: 2026-10-24 Event date: 2025-12-11 Tags: [Ai2](https://rangoon.ai/insights/tags/ai2/), [Molmo 2](https://rangoon.ai/insights/tags/molmo-2/), [multimodal AI](https://rangoon.ai/insights/tags/multimodal-ai/), [open models](https://rangoon.ai/insights/tags/open-models/) Source note: Historical reporting: Ai2 released Molmo 2 on December 11, 2025 as an open video-and-multi-image model family. The launch names 4B, 8B, and Olmo-based 7B variants, new datasets, and research-and-education use. The official code repository lists released pretraining, SFT, and long-context checkpoints. Ai2’s benchmark and peer comparisons are publisher results, not independent validation or a claim that every training stage was open on launch day. Image: A video question is tied to selected frames, a Molmo 2 checkpoint, and spatial-temporal evidence. Original Rangoon explanatory diagram: Source boundary: dated primary release reporting; original editorial analysis. No vendor interface, photograph, measurement result, or endorsement implied. ## Key takeaways - Molmo 2 separated 4B, 8B, and fully open 7B variants by capability and openness. - Video pointing and tracking require temporal evidence, not only a fluent caption. - Published checkpoints make artifact provenance inspectable without proving local performance. ## Ai2 released Molmo 2 on December 11, 2025 [source 1](https://allenai.org/blog/molmo2) Ai2 announced Molmo 2 on December 11, 2025 as a model family for single images, multi-image inputs, and video. The release names an 8B Qwen 3-based model, a 4B efficiency-oriented variant, and Molmo 2-O at 7B using an Olmo language backbone. It also reports new datasets and evaluation tools for grounded video work. This is a dated release of model variants and supporting artifacts. It does not establish that a particular video system can identify an event correctly, process every clip length, or meet a required latency or safety target. Those conditions depend on the selected model, frames, and application pipeline. ## The Olmo-based 7B variant makes end-to-end openness a separate property from model size [source 1](https://allenai.org/blog/molmo2) Ai2 describes Molmo 2-O at 7B as a fully open end-to-end option, including the language model, vision encoder, and training checkpoints. The 4B and 8B variants use Qwen 3-based language components. This is an important release distinction: “Molmo 2” is a family name, while openness and underlying components vary by variant. A user should identify the exact model rather than borrowing a property from a sibling checkpoint. The blog also states research and educational use under Ai2’s guidelines, so availability does not settle a downstream organization’s use decision. Weights, datasets, code, and permitted use should each be checked from their applicable record. ## The official repository exposes checkpoints at pretraining, SFT and long-context stages [source 2](https://github.com/allenai/molmo2) Ai2’s Molmo2 repository lists pretraining, supervised fine-tuning, and long-context SFT checkpoints for the 4B, 8B, and O-7B variants. That artifact structure helps a researcher distinguish a base-style checkpoint from a post-trained checkpoint. It also creates a reproducibility obligation: a result reported with long-context SFT should not be attributed to a pretraining artifact, and an evaluation should record the checkpoint identifier. The repository includes training and configuration material, but a published implementation is not a claim that a local run matched the publisher’s data pipeline or compute environment. It provides inspection and starting points, not automatic replication. ## Ai2’s video-grounding claims need temporal fixtures and publisher-method boundaries [source 1](https://allenai.org/blog/molmo2) [source 2](https://github.com/allenai/molmo2) Ai2 reports video pointing, counting, tracking, and question-answering comparisons for Molmo 2, including performance against open and proprietary peers. The release’s key contribution is not merely that it can summarize video; it says the model can return spatial and temporal grounding evidence. That claim should be tested with fixtures that identify the required frame range, object identity, occlusion, re-entry, and acceptable uncertainty. A model can produce a plausible count while pointing to the wrong time or object. Ai2’s benchmark results are meaningful source material under its stated methods, but they do not independently validate a security camera, industrial feed, or user-uploaded clip. Local evaluation should retain clip provenance, sampling policy, checkpoint, and expected evidence locations. ## Sources 1. [Ai2: Molmo 2 launch](https://allenai.org/blog/molmo2) 2. [Ai2: Molmo2 official code and checkpoints](https://github.com/allenai/molmo2) ## Related reading - [Hugging Face released SmolVLM2 as a compact family for image and video understanding](https://rangoon.ai/insights/archive-hugging-face-smolvlm2-video-february-2025/) - [Ai2 released olmOCR as an open 7B vision model for PDF extraction](https://rangoon.ai/insights/archive-ai2-olmocr-pdf-extraction-february-2025/) - [OLMo 3 made end-to-end openness a versioned supply chain](https://rangoon.ai/insights/archive-olmo-3-release-2025/) ## More information - [Documentation index](https://rangoon.ai/llms.txt) - [AI access and policies](https://rangoon.ai/ai/) - [Source repository](https://github.com/hypler-dev/rangoon)