olmOCR 2

By our estimate, olmOCR 2 needs about 6.3 GB at Q4, 10.6 GB at Q8 and 19.6 GB at FP16 of VRAM. That covers the weights plus headroom for a short context, and longer contexts need more.

Yes, by our estimates olmOCR 2 fits on an 8 GB GPU at Q4 (about 6.3 GB) and at higher precision on larger cards.

Yes, a Mac with at least 16 GB of unified memory should run olmOCR 2 at Q4, which needs an estimated 6.3 GB. This assumes macOS gives the GPU about 75% of unified memory.

Yes, olmOCR 2 has open weights, so you can run it on your own hardware and keep prompts on your own machines.

Run olmOCR 2 locally

Yes, olmOCR 2 accepts images as input alongside text.

Similar models from other labs include MiniMax-M2 (MiniMax), Llama 3.1 Nemotron Safety Guard 8B v3 (NVIDIA) and Llama Nemotron Embed 1B v2 (NVIDIA). They are the closest matches by kind and release date.

MiniMax-M2olmOCR 2 vs MiniMax-M2Llama 3.1 Nemotron Safety Guard 8B v3olmOCR 2 vs Llama 3.1 Nemotron Safety Guard 8B v3Llama Nemotron Embed 1B v2olmOCR 2 vs Llama Nemotron Embed 1B v2

olmOCR 2 has open weights under the Apache 2.0 license, so you can download and run it yourself.

Yes, olmOCR 2 is released under the Apache 2.0 license, which allows commercial use. Check the license text for attribution terms.

The olmOCR 2 weights are on Hugging Face at allenai/olmOCR-2-7B-1025.

Run olmOCR 2 locally

olmOCR 2 was made by Ai2. It was released on October 22, 2025.

More from Ai2

Yes, olmOCR 2 has open weights, so you can run it on your own hardware without API fees.