What is Molmo2-8B?
Molmo2 was released by the Allen Institute for AI (Ai2) on 11 December 2025, in 4B and 8B versions built on Qwen3 and a 7B version built on Ai2's own Olmo. The 8B model uses Qwen3-8B with a SigLIP 2 vision encoder and answers questions about images, several images or video, and can point to and track objects across video frames.
The weights are under Apache 2.0, and Ai2 published the nine new datasets behind it, more than 9 million examples, none generated by closed models; some are licensed for non-commercial research only. The full training code followed in March 2026. Ai2 describes the model as intended for research and educational use. It is still the newest model in the Molmo line.
Is Molmo2-8B worth it?
On Ai2's tests, Molmo2-8B is the strongest open model for grounding answers in video: pointing at an object, counting it and following it across frames. In Ai2's paper it beats Gemini 3 Pro on video pointing, 38.4 against 20.0 F1, and on video tracking, 56.2 against 41.1 J&F, and it leads Qwen3-VL on video counting, 35.5 against 29.6. The paper notes that even closed models mostly lack this kind of grounding.
It is also unusually open. Ai2 published the datasets, the code and the base model behind the 7B Olmo variant, and trained on less than one-eighth of the video Meta used for PerceptionLM, 9.19 million videos against 72.5 million. For a team that has to explain where a model's abilities came from, that matters as much as the scores.
The limits are on reasoning and length. Ai2 says it lags slightly on open-weight reasoning benchmarks, larger closed systems keep a clear lead on long-video questions, and it samples at most 128 frames at up to 2 frames per second, so a ten-minute clip is seen as scattered stills. Ai2's own blog notes that no model, Molmo2 included, reaches 40% accuracy on video grounding.
At about 8.7 billion parameters it fits on one 24GB GPU in BF16. It is ageing: the model is nearly a year old and has 44,427 downloads in the last month, far below newer general models such as Qwen 3.8, which now read video too.
Where it earns its place
A sports-analytics start-up
Count and follow players or the ball through short broadcast clips
Ask it to point to or track named objects and read back per-frame coordinates. One 24GB GPU runs it; clips longer than 128 sampled frames need splitting first.
A warehouse robotics team
Tell a robot where the item named in an instruction is in a camera feed
Molmo2's pointing output gives pixel coordinates the controller can use. Check the dataset licences before commercial deployment, because some training sets are non-commercial.
An academic lab studying how VLMs learn from data
Retrain a video model from scratch and change the data mix
The released datasets, code and fully open Molmo2-O 7B variant make that possible without reverse-engineering, which few rival models allow.
Before you commit
- Setup
- Install transformers 4.57.1, torch, decord2 and molmo_utils, then load with AutoModelForImageTextToText and trust_remote_code; Ai2's examples cover video QA, pointing and tracking.
- Integrations
- Hugging Face Transformers with remote code, Ai2 Playground, the open allenai/molmo2 training code; no paid hosted API listed on OpenRouter.
- Data and privacy
- Apache 2.0 weights; self-hosted, video stays local. Some training datasets are for non-commercial research only, and Ai2 says the model is intended for research and educational use.
- Cost at scale
- Self-hosted only: about 17GB of BF16 weights, so one 24GB GPU per instance.
- Limits
- Up to 128 frames at 2 fps or less per clip; behind larger closed models on long video; weaker on reasoning benchmarks; no Molmo3 announced.
What’s good
- Points at and tracks objects in video; 56.2 J&F tracking against Gemini 3 Pro's 41.1
- Apache 2.0 weights with published datasets and training code
- Fits on one 24GB GPU
- Trained without data generated by closed models
What’s not
- Lags on reasoning benchmarks and long-video questions
- At most 128 frames per clip
- Some training datasets are non-commercial; Ai2 frames it as a research model
- No paid hosted API
Best for
Researchers and robotics or video-analytics teams that need open, explainable video grounding they can retrain.
Molmo2-8B alternatives
Molmo2-8B: common questions
- Is Molmo2-8B worth it?
- Molmo2-8B is the open model to reach for when you need to know where in a video something is, not just whether it is there. We rate it 3.9 out of 5.
- How much does Molmo2-8B cost?
- Free weights under Apache 2.0; some training datasets are non-commercial. No paid hosted API listed.
- What are the downsides of Molmo2-8B?
- Lags on reasoning benchmarks and long-video questions; at most 128 frames per clip; some training datasets are non-commercial; Ai2 frames it as a research model; no paid hosted API.
- Who is Molmo2-8B best for?
- Researchers and robotics or video-analytics teams that need open, explainable video grounding they can retrain.
- How much does Molmo2-8B cost for a team or in production?
- Self-hosted only: about 17GB of BF16 weights, so one 24GB GPU per instance.
- Is Molmo2-8B safe for company data?
- Apache 2.0 weights; self-hosted, video stays local. Some training datasets are for non-commercial research only, and Ai2 says the model is intended for research and educational use.
- What are the best Molmo2-8B alternatives?
- The closest alternatives we have reviewed are SAM 3 (Segment Anything Model 3), LocateAnything-3B, MiniCPM-V 4.7 (35B-A3B and 1B) and Qwen3.8-27B.
Answers checked Nov 7, 2026.