Molmo 2

What it is

Molmo 2 is the second generation of Ai2's family of open vision-language models, announced on 11 December 2025. The earlier Molmo worked on still images; Molmo 2 extends the line to video^[1]. A vision-language model combines a visual encoder with a language model so that the system can answer questions about what it is shown. What marks out the Molmo line is that its answers can be tied to places in the input — a point, a box, a timestamp — rather than only to a sentence of description.

Where it comes from

Ai2 said that the original Molmo set a new state of the art on static-image benchmarks in 2024 and pioneered image pointing^[2]. The organization also reported that Molmo had been downloaded millions of times, and that the PixMo dataset released with it became a go-to resource for teams replacing larger proprietary models^[2]. Molmo 2 is the follow-up to that work and keeps the open-release posture of the first generation.

The three models

The release contains three models. Molmo 2 8B and Molmo 2 4B are built on Qwen 3, while Molmo 2-O 7B is built on Olmo, Ai2's own model family, which the organization describes as giving a fully open model flow^[1]. The sizes are small by the standards of frontier systems: the largest of the three has 8 billion parameters, and the point of the family is that the weights are published rather than held back.

What it does with video

For video, the listed abilities are pointing, multi-object tracking with persistent IDs, dense captioning, and grounded question answering that returns timestamps and locations^[1]. These are related skills. Pointing marks a specific spot in a frame. Tracking follows several objects through a clip while giving each a persistent ID, so that the same identifier refers to the same object from the beginning of the clip to the end. Dense captioning describes the clip in detail rather than in a single sentence. Grounded question answering answers a question and says when and where in the clip the answer occurs^[1].

The model samples up to 128 frames per clip, at up to 2 frames per second^[1]. That framing shapes what the model sees: a long clip is read by taking frames at intervals rather than by watching every frame, so the effective resolution in time depends on how long the clip is.

Training data

Ai2 trained Molmo 2 on about 9.19 million videos. The same announcement gives 72.5 million videos for Meta's PerceptionLM^[1], so the video corpus behind Molmo 2 is a good deal smaller than that comparison point.

Reported results

By Ai2's own account, Molmo 2 8B beats the original Molmo 72B on image pointing and grounding, meaning a smaller model outperforms the earlier flagship on those tasks^[1]. Ai2 also said the model beats Gemini 3 Pro by a wide margin on video tracking^[1]. On an average across 11 image benchmarks, Ai2 placed Molmo 2 first among open-weight models, just behind GPT-5 and GPT-5 mini^[1]. These are the organization's own comparisons, reported alongside the release.

Licensing

The weights are released under the Apache 2.0 licence^[1]. Some third-party training datasets are restricted to non-commercial research^[1], so the model weights are permissive even though not every piece of data behind them can be reused freely.

Where it sits

Molmo 2 arrives at a moment when open-weight vision-language models are compared directly with closed frontier systems, and Ai2's framing of the release is built around exactly that comparison: first among open-weight models on an 11-benchmark image average, just behind GPT-5 and GPT-5 mini^[1]. The move to video, with object tracking and answers that carry timestamps and locations, shifts the family from describing a single picture towards saying where and when something happened inside a clip^[1]. The earlier generation's record — a state-of-the-art claim on static images in 2024, pioneering image pointing, millions of downloads and a dataset that other teams adopted^[2] — is the baseline the second generation is measured against.