Molmo2-4B

Molmo2-4B

Answers questions about images, videos and groups of images, and can point to what it finds. An open vision-language model from Ai2 for developers and researchers who run models themselves.

Made by allenai on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: Molmo2-4B is an open vision-language model from the Allen Institute for AI (Ai2). It reads images, videos and sets of several images, answers questions about them in text, and can point to things it finds. It sits on Qwen3-4B-Instruct and uses SigLIP 2 for the vision side.

Who it suits: It suits developers and researchers who want to run an open model for image and video questions on their own machine. It handles English only, according to the record.

Industry | Who uses it | How they use it | Example
Research | Machine learning researchers | Test video question answering and pointing on their own clips | Ask which animal appears in a short video
Media | Video archive teams | Caption and count things in clips | Count the objects in a batch of short clips
Software | App developers | Add image and video understanding to a product | Write captions for uploaded photos

People who only need plain text chat can skip it, since this model is built around pictures and video.

Where to use it:

• Python with the transformers library, on Windows, macOS or Linux
• A conda environment, which is how the model card sets it up
• Any machine that can hold the weights, about 19.4 GB in F32 or an estimated 2.4 GB at 4-bit

Use case | Who | What they get back
Video question answering | Developers | A written answer about what happens in a clip
Pointing in video | Researchers | Coordinates for the things you asked about, as text you can parse
Captioning and counting | Content teams | Captions and counts for images and short videos

How to set it up: You need a few things first.

• Python 3.11, which the card installs through conda
• Enough disk space and memory for the weights
• The Hugging Face hub tool, if you want the files on disk first

1. Download the repository files with the hub tool.
2. Create and activate a conda environment with Python 3.11.
3. Install transformers 4.57.1 and the other packages the card lists.
4. Load the processor and the model with trust_remote_code turned on.

pip install -U huggingface_hub
hf download allenai/Molmo2-4B

conda create --name transformers4571 python=3.11
conda activate transformers4571
pip install transformers==4.57.1
pip install torch pillow einops torchvision accelerate decord2 molmo_utils

How to use it: You write a message that holds a text question and a video or image. The processor turns it into model input, the model generates tokens, and you decode them back to text. The card has a general video question example and a pointing example.

from transformers import AutoProcessor, AutoModelForImageTextToText
import torch

model_id="allenai/Molmo2-4B"

Tool | What it does
AutoProcessor | Prepares the text and video for the model
AutoModelForImageTextToText | Loads Molmo2-4B and generates the answer
molmo_utils | A helper package whose process_vision_info is used in the pointing example
General Video QA | Asks a free-form question about a video
Pointing Video QA | Asks the model to point at things and turns the reply into coordinates

Example: Attach a short video and ask "Which animal appears in the video?", the question the card itself uses. The model replies with a sentence naming the animal it sees, and you print that text after decoding the generated tokens.

For newcomers and experts: A vision-language model looks at pictures or video and answers in words. For a first try, set up the environment, run the general video example with a short clip, and change the question to something about your own footage.

Underneath, a language model from the Qwen3 family is paired with a SigLIP 2 vision encoder. Pointing answers come back as text with coordinates scaled by 1000, which the example code converts to pixels. The weights ship in F32, so memory is the main cost, and trust_remote_code means you run code that comes with the repository.

Video or image plus a text question
Processor builds the model input
Model generates answer tokens
Tokens are decoded back to text

Good to know:

• License: Apache 2.0.
• Cost to run: no API key or paid service named. You supply the compute, and the weights download from Hugging Face.
• Data it reaches: the images and videos you pass in. The card's sample loads a video from the web, and trust_remote_code runs code from the repository.
• Not for: languages other than English, since the record lists only en. Its training code and evaluations are promised for a later date, so they are not part of this download.

On Hugging Face, 13 November 2026

• Task: image text to text
• Library: transformers
• Parameters: 4.9B
• Weights: about 19.4 GB in F32; about 2.4 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 26,915
• Likes: 57
• License: apache-2.0
• Languages: en
• Based on: google/siglip-so400m-patch14-384
• Quantized versions: 6
• Fine-tunes: 15
• Tags: transformers, safetensors, molmo2, image-text-to-text, multimodal, olmo, molmo, conversational

Links

• Model: https://huggingface.co/allenai/Molmo2-4B
• Files: https://huggingface.co/allenai/Molmo2-4B/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download allenai/Molmo2-4B

Then follow the usage in its model card. Nothing in this kit runs by itself.