Survey of Open Source Vision Language Models (2026)
Younes El Hjouji
How much of your time do you spend with your eyes open? The question seems ridiculous, we do not even think of seeing as an activity. In any given day, we will decide to say many things, and read quite a few, but through it all we are constantly watching the world around us.
This ratio is inverted for an LLM, the vast majority of an LLM’s training data and interactions are text based, with the occasional image or video thrown in. This is not because interacting with the world through text is superior, but because text was the easiest avenue for general AI to enter into the world. Much like the newspaper dominated before the advent of the Television, we at Overshoot believe this is the age of text, to be followed by the age of vision.
At this stage, vision in LLMs is added on after the fact. Images and videos are pushed through vision encoders, converted to tokens that neatly fit alongside text input tokens for an LLM’s ingestion. This approach (primitive though it may be) is generating impressive results for visual understanding and is changing the computer vision landscape for security, home monitoring, blind assistance, robotics, and many more.
Some of the best vision models available are open source models, with the Qwen3.5 and Qwen3-VL families at the forefront of model performance and latency. Since summer 2025, the range of possible applications has expanded to include realtime general vision applications. With response times as low as 200ms, rivaling human reaction times (try it for yourself in playground.overshoot.ai).
The following is a list of the most important open source video language models as of March 2026. We aim to encompass all significant vision language models that are published later than December 2024 and which support text, image, and importantly video input.
The Qwen3.5 family is the first major model release where every model is natively multimodal via early fusion training. There is no separate “-VL” variant. Every model handles text, images, and video from a single set of weights. See our detailed breakdown for benchmarks and model selection guidance.
