TinyGPT-V

TinyGPT-V is an open-source multimodal large language model published on GitHub that achieves strong vision-language performance with a small language-model backbone, designed to train and run on modest hardware. The project has gathered around 1,300 GitHub stars and is still updated in 2026.

Overview

Most multimodal LLMs stack a huge language model behind a vision encoder, putting them out of reach for small labs. TinyGPT-V instead builds on the compact Phi-2 backbone and adds carefully designed visual projection and quantisation-aware training stages, reaching competitive scores on benchmarks such as VQAv2 and GQA while training on a single 24 GB GPU and inferring with far less memory than comparable systems. The repository provides the official implementation of the paper, training recipes across its staged alignment pipeline, evaluation scripts and model checkpoints. It demonstrates that efficient architecture and training choices can substitute for raw scale in multimodal tasks like image question answering, captioning and visual reasoning — a useful blueprint for anyone building vision-language systems under real compute budgets.

Key Features

• Multimodal LLM built on a small Phi-2 backbone
• Competitive VQA and visual-reasoning benchmark results
• Trains on a single 24 GB GPU with quantisation-aware stages
• Low-memory inference for edge and workstation deployment
• Full training pipeline and evaluation scripts included
• Released model checkpoints ready to download

💳 Pricing

TinyGPT-V is free and open source under the BSD 3-Clause license. Costs are limited to your own training and inference compute.

Who Is It For

Researchers and developers who want a capable vision-language model they can actually train and serve without a GPU cluster.