PaliGemma – Google’s Cutting-Edge Open Vision Language Model

Google released PaliGemma, a 3-7B vision-language model that takes a single image and text prompt, then outputs text. It runs on a single T4 GPU and beats Flamingo-80B on VQAv2 while training on 1/100th the data.

Most teams still pipe images through GPT-4V at $0.015 per 512×512 token; PaliGemma inference on Vertex costs ~$0.0003. Running a 24-frame video summary with GPT-4V burns $2.88; PaliGemma does it for five cents. Stop paying OpenAI for OCR and captioning work.

Build MVPs or batch jobs that need cheap, fast vision-to-text. Skip if you're doing real-time agents that need tool calling or reasoning.