Can I run DeepSeek-OCR?

DeepSeek-OCR is a specialist, not a chat model: it reads images and turns documents, screenshots, and scanned pages into text. It is a mixture-of-experts design, so even though the full model is 3B parameters, only about 0.57B of them activate for any given token. That keeps it quick, but the catch with MoE is that you still need room for the whole thing in memory, not just the active slice. At a 4-bit quant it lands around 1.8 GB and wants roughly 4 GB of RAM minimum, so it sits comfortably on a 12 GB RTX 3060, a small Apple Silicon Mac, or even a modern CPU box.

In practice it feels less like waiting on a chatbot and more like running a fast batch job. On an RTX 3060 12GB you can expect somewhere around 886 tokens per second, an M-series Max pushes past 1000, and a 4090 is in another league at roughly 2479 tok/s. Even pure CPU on DDR5 manages about 148 tok/s, which is usable for occasional pages. The context window is a modest 8K, which is plenty for OCR work since you are feeding it one image at a time; at full context the whole thing tops out near 2.5 GB of memory, so headroom is rarely the problem here.

Do not reach for this as a general assistant. If you want reasoning or step-by-step answers, a sibling like DeepSeek R1 7B generally serves you far better, and Ministral 3 3B is the more sensible pick when you want vision plus actual conversation in one model. DeepSeek-OCR's standout trait is narrow and real: it is a tiny, fast, document-to-text engine that runs almost anywhere. It ships under a plain MIT license, so you can use it freely in production and commercial work without the provider-specific strings that come attached to many open-weight releases.