Read this before you buy
DocuAI runs every AI model on your own machine — nothing is offloaded to a server — so the hardware you run it on decides how fast it feels. This page is the honest version: what works, what is slow, and what is not supported.
Quick answer: a Windows 10/11 PC with an NVIDIA graphics card (6 GB of video memory or more), 16 GB of RAM and 25 GB of free disk runs everything well. Anything less still works, with the trade-offs below.
Operating system
| System | Status |
|---|---|
| Windows 11 (64-bit) | Recommended |
| Windows 10 (64-bit) | Supported (minimum) |
| Windows on ARM, 32-bit Windows | Not supported |
| macOS, Linux | Not supported in 1.0 (on the roadmap) |
DocuAI also needs the Microsoft Visual C++ Redistributable 2022 (x64); the installer normally has it already.
Graphics card — what you get at each size
DocuAI detects your graphics card at startup (Settings → Hardware shows the result) and sizes its models to the video memory it finds.
| Video memory | What you get |
|---|---|
| 12 GB and up RTX 3060 12 GB, RTX 4070, RTX 4080/4090, … | Everything loaded at once: chat, vision classifier, structured OCR. Longest chat context. Scans and chat run side by side without reloads. 16 GB and up keeps every model resident permanently. |
| 8 GB RTX 3070, RTX 4060, RTX 3060 8 GB, … | Full feature set, extended chat context. Scan-only models are unloaded after 10 minutes idle and reload on the next scan (5–10 s); the chat model is released after 10 minutes with no question or scan and reloads on the next question (60–90 s, shown in the chat panel). |
| 6 GB RTX 2060, GTX 1660 Ti, RTX 3050, … | Full feature set, standard chat context. Scan-only models are dropped after 2 minutes idle and the vision model loads on first scan rather than at startup; the chat model is released after 2 minutes with no question or scan and reloads on the next question (60–90 s). Expect short pauses when switching between scanning and chat. |
| Under 6 GB, or no NVIDIA card | Chat falls back to a smaller model, or to the CPU. See below. |
NVIDIA (CUDA 12 or newer driver) is the recommended and best-tested path. You only need a current NVIDIA driver; no separate CUDA install.
AMD, Intel Arc and integrated graphics run through the bundled Vulkan runtime. It works and accelerates chat and vision, but it is less tested and typically slower on the same class of card.
CPU only: every feature still runs, but chat generates at roughly one second per word (a paragraph is a minute or more) and vision-based classification of scanned images is slow. Search, text extraction, Key Fields and the Transactions and Spend views are unaffected; they do not depend on the chat model.
Memory and disk
- 16 GB of RAM recommended. Models that do not fit in video memory are held in system RAM instead, so a smaller graphics card means a larger RAM footprint. 8 GB is the floor; other heavy applications will compete.
- Download: 10.6 GB. Plus about 10 GB of working space for the encrypted index — plan for 25 GB free.
- An SSD is strongly recommended: first-load times on a hard disk are several times longer.
- DocuAI needs at least 500 MB free on its own drive at all times or it will refuse to start.
What to expect on first use
- The first chat answer after launch takes 60–90 seconds while the chat model loads (longer on CPU or a hard disk). The chat panel shows a “warming up” notice; later answers start within a second or two on a graphics card.
- The first scan after launch waits for the reading and vision models to warm up; the status bar labels this as warm-up rather than progress.
- Indexing a large library takes time: digital PDFs are fast; scanned image pages take many times longer, and longer again on CPU. Keyword search is available as soon as a document is imported.