What Are Vision-Language Models?

A vision-language model (VLM) takes an image and a text prompt as input and produces a text response. It can answer questions about images, locate objects described in natural language, and follow instructions that reference visual content. Datature Vi lets you fine-tune VLMs on your own data so the model learns your specific domain and task.

How did we get here? From classification to VLMs

Computer vision has evolved through three major phases. Each phase gave models more flexibility in how they understand and describe images.

Image classification (2012-2015) was the first breakthrough. Given an image, the model picks one label from a fixed list: "cat," "dog," "car." The model can only choose categories it was trained on. Adding a new category means retraining from scratch.

Object detection (2015-2020) added spatial awareness. The model draws bounding boxes around objects and labels each one. Still limited to pre-defined categories, but now the model can find multiple objects and say where they are. This is where formats like YOLO and COCO became standard.

Vision-language models (2021-present) removed the fixed-category constraint. Instead of picking from a list, VLMs generate free-form text. You can ask a VLM "find the dented can on the second shelf" and it will locate it, even if "dented can" was never a training category. The input is flexible (any text prompt) and the output is flexible (any text response, with optional bounding boxes).

How do VLMs work?

A VLM has three main parts that work together. Traditional CV models have eyes but no language comprehension. VLMs have both.

Key concepts you'll encounter

These are the terms you'll see throughout the docs. Each expands with a plain-language explanation. You don't need to understand all of them before starting the quickstart.

Which VLM architectures does Datature Vi support?

Datature Vi supports seven model architectures across three Qwen generations plus specialist models. Start with Qwen3.5 4B to validate your setup, then scale to Qwen3.5 9B for production.

For full architecture details, hardware requirements, and benchmark citation pointers, see Model Architectures.

How do I choose a model size?

Model size (parameter count) is not "bigger is better." The right choice depends on your dataset size, task complexity, and deployment constraints.

A 27B model trained on 50 images will often perform worse than a 4B model on the same data. Large models have so many parameters that they can memorize small datasets instead of learning general patterns. Match model size to data size first, then scale up if evaluation metrics plateau.

For more on how model size interacts with training settings, see How Does VLM Training Work?

Frequently asked questions

Further reading

• How to Fine-Tune Qwen2.5-VL: A step-by-step walkthrough of fine-tuning one of Datature Vi's supported architectures.
• VLM Evaluation Metrics Guide: How to measure and interpret VLM performance across tasks.
• LoRA glossary: Quick reference for Low-Rank Adaptation.
• Fine-tuning glossary: Quick reference for the fine-tuning process.

Related resources