Tasks and inputs
Molmo2-4B is listed for Image-text-to-text, Video-text-to-text. It accepts image, video, text and produces text. Use the publisher’s input format and preprocessing for your chosen deployment.
Molmo2-4B is Ai2's open 4B vision-language model for video, multi-image and image understanding, with pointing and object tracking built in.
The specification score summarizes recorded features and limits. It is not a benchmark result.
Pricing conditions: Open weights for self-hosting; no hosted API price. Hardware cost is yours.
Recorded modality support shows what this model can accept and produce. An undocumented modality is shown as unknown rather than assumed to be unsupported.
Accepts
Produces
Ask most small vision-language models where something is in a video and they will describe it. Molmo2-4B will point at it, frame by frame, and can track it. That is the difference. Grounding, carried over from the original Molmo's image pointing and extended to video and multi-image inputs, is what separates Ai2's second-generation family from the many small models that can only talk about what they see.
Ai2 announced Molmo 2 on 11 December 2025; the 4B weights went up a few days later. It is the efficiency-focused option of three, next to an 8B model and a fully open 7B built on Olmo.
Two parts. The language model is Qwen3-4B-Instruct-2507, and the vision backbone is SigLIP 2 (so400m, patch 14, 384px). The family was trained on Ai2's own curated image-text and video-text collections plus public third-party data, 9.19 million videos in all. Ai2 contrasts that with the 72.5 million used by Meta's PerceptionLM.
Capabilities exposed in the examples:
Ai2 reports an average of 62.8 across 15 academic video benchmarks for the 4B. On the same table:
| Model | Average |
|---|---|
| GPT-5 | 70.6 |
| Gemini 2.5 Pro | 71.2 |
| GPT-5 mini | 65.0 |
| Molmo2-8B | 63.1 |
| Molmo2-4B | 62.8 |
| Eagle2.5-8B | 60.7 |
| Qwen3-VL-4B | 58.1 |
| InternVL3.5-4B | 53.4 |
Ai2 also says the 4B outperforms open-weight and open-data peers on short videos, counting and captioning, while staying competitive on long videos.
Setup is short.
The quick start pins Transformers 4.57.1 and adds einops, decord2 and molmo_utils, loading with trust_remote_code. Since March 2026 the full training codebase is open too, with pretraining, SFT and long-context SFT scripts and vLLM inference support.
Read this part.
The weights are Apache-2.0, intended for research and education under Ai2's responsible-use guidelines. The training data include third-party sets restricted to academic and non-commercial research, so check whether that matters for your use before shipping a product.
Counting and localisation in short clips, which is where Ai2 claims the clearest lead over peers of the same size and where a wrong answer is easy to verify by hand. Give it fifty of your own videos with hand-counted objects and compare the counts and point accuracy with your current model. Other video-text models and image-text models are listed in the directory, or line candidates up on the compare page.
Compare the model on the work you need to complete.
Molmo2-4B is listed for Image-text-to-text, Video-text-to-text. It accepts image, video, text and produces text. Use the publisher’s input format and preprocessing for your chosen deployment.
Measure a completed task using the prompts, documents and tools your application will use. Include reasoning and retries in the result.
The weights are downloadable under Apache-2.0. Check runtime support and size the complete model and working context for your expected concurrency.
These records share a directory category. Check the source card, output type and deployment requirements before treating a model as a replacement.
Tencent
WeMM-Video is Tencent's open Chinese-English omni-modal family (2B to 35B-A3B) that reasons over video frames and their soundtrack and answers in text.
NVIDIA
Flora-Captioner is NVIDIA's 27B non-commercial video model that writes one action instruction for a chosen interval of a robot manipulation video.
NVIDIA
Flora-Rewarder is NVIDIA's 27B non-commercial reward model that scores robot episodes for success and grades each half-second of action quality.
perceptron
Perceptron Mk1.5 is an embodied reasoning model for robots, drones and glasses: video, image and audio in; text plus boxes, points and tracks out.
Ai2 (Allen Institute for AI)
allenai/MolmoPoint-Vid-4B is a source-linked Hugging Face repository assigned to Video-text-to-text. Its recorded task interface accepts video, text and produces text.
chenjoya
chenjoya/videollm-online-8b-v1plus is a source-linked Hugging Face repository assigned to Video-text-to-text. Its recorded task interface accepts video, text and produces text.