PaddleOCR-VL

By our estimate, PaddleOCR-VL needs about 1.2 GB at Q4, 1.7 GB at Q8 and 2.7 GB at FP16 of VRAM. That covers the weights plus headroom for a short context, and longer contexts need more.

Yes, by our estimates PaddleOCR-VL fits on an 8 GB GPU at FP16 (about 2.7 GB) and at higher precision on larger cards.

Yes, a Mac with at least 8 GB of unified memory should run PaddleOCR-VL at Q4, which needs an estimated 1.2 GB. This assumes macOS gives the GPU about 75% of unified memory.

Yes, PaddleOCR-VL has open weights, so you can run it on your own hardware and keep prompts on your own machines.

Run PaddleOCR-VL locally

Yes, PaddleOCR-VL accepts images as input alongside text.

Similar models from other labs include Llama Nemotron Embed 1B v2 (NVIDIA), PaddleOCR VL (Paddlepaddle) and olmOCR 2 (Ai2). They are the closest matches by kind and release date.

Llama Nemotron Embed 1B v2PaddleOCR-VL vs Llama Nemotron Embed 1B v2PaddleOCR VLPaddleOCR-VL vs PaddleOCR VLolmOCR 2PaddleOCR-VL vs olmOCR 2

PaddleOCR-VL has open weights under the Apache 2.0 license, so you can download and run it yourself.

Yes, PaddleOCR-VL is released under the Apache 2.0 license, which allows commercial use. Check the license text for attribution terms.

The PaddleOCR-VL weights are on Hugging Face at PaddlePaddle/PaddleOCR-VL.

Run PaddleOCR-VL locally

PaddleOCR-VL was made by Baidu. It was released on October 16, 2025.

More from Baidu

Yes, PaddleOCR-VL has open weights, so you can run it on your own hardware without API fees.