Molmo 2 (4B, 8B and Molmo2-O 7B)

2025-12-16 · Publisher : Ai2 · Domain : Artificial Intelligence

Ai2 releases Molmo 2, a family of open multimodal models for video, multiple images and pointing, in 4B and 8B versions (Qwen 3 base) and Molmo2-O 7B (Olmo base).

### Context Molmo (2024) had introduced image pointing. Molmo 2 extends this capability to video: pointing, object tracking with persistent identifiers, and localization in space and time. Ai2 compares it notably with Gemini 3 Pro and GPT-5 in its announcement. ### What the model brings - Video pointing, object tracking with persistent identifiers, dense video captioning, counting by pointing and spatio-temporal grounding. - Nine new datasets (9.19 million examples) including Molmo2-Cap (104,000 videos), Molmo2-VideoPoint (more than 300,000 queries) and Molmo2-VideoTrack (3,600 clips, 15,000 queries). - On video tracking, Ai2 announces that Molmo 2 outperforms the other open models and beats Gemini 3 Pro “by a wide margin”. - On the average of 11 image benchmarks, the 8B ranks just behind GPT-5 and GPT-5 mini and ahead of Gemini 2.5 Pro, according to Ai2. - Release of the weights, the training and inference code, the datasets and the evaluation tools. ### Architecture and specifications Three models: Molmo 2 8B and 4B (Qwen 3 base) and Molmo2-O 7B (Olmo base, fully open chain). Inputs: images, image series, videos of variable length. License: Apache 2.0, with the caveat that some third-party datasets used in training are subject to academic research or non-commercial restrictions. Openness: weights, code, data and evaluations; only the Molmo2-O variant is built on a base that is itself fully open, and the Qwen 3 variants inherit Qwen's limits. Available on Hugging Face, GitHub, the Ai2 Playground and the OpenRouter API. ### Cost and access | Item | Detail | |---|---| | Official API | No Ai2 API; access via OpenRouter (price not recorded) | | Hardware | Not disclosed (4B and 8B); about 17 GB in BF16 for the 8B (calculated on 8.66 billion parameters) | | Weights | Molmo2-4B, Molmo2-8B, Molmo2-O-7B on Hugging Face (December 14, 2025) | ### Benchmarks | Comparison published by Ai2 (qualitative) | Result | |---|---| | Video tracking | Ahead of open models and ahead of Gemini 3 Pro “by a wide margin” | | Reasoning on images (average of 11 benchmarks) | Molmo 2 8B just behind GPT-5 and GPT-5 mini, ahead of Gemini 2.5 Pro | | Human preference | Top among open models | Source: Ai2 blog post and press release of December 16, 2025, vendor comparisons; the numerical scores do not appear in the text consulted. ### Versus closed models This is the only milestone in this series where the vendor directly compares its model with closed models: Gemini 3 Pro on video tracking and GPT-5 and GPT-5 mini on the average of the image benchmarks. The figures could not be recorded, so these gaps remain Ai2 claims. Ai2 chooses Gemini 3 Pro and GPT-5 for comparison, recent closed models from Google and OpenAI at that date, which indicates the level aimed at, without the gap being quantified here. The 8B model takes up about 17 GB in BF16 and can be self-hosted, whereas closed models are API services. The open versus closed choice therefore comes down to control over video data and inference cost, not general parity. ### Key takeaways Molmo 2 brings pointing and video tracking capabilities to open models, with a fully open Molmo2-O variant. The comparisons with closed models are Ai2's.

Tags : LLM, Multimodal, Vidéo, Open Source, Entièrement ouvert (Molmo2-O), Ai2, Molmo, Apache 2.0