Quick summary: Running AI locally is limited mainly by GPU memory (VRAM). A single NVIDIA card with 16 GB of VRAM, or a Mac with Apple Silicon, is enough to start with a capable model for chat and document Q&A. Bigger models and more users need more VRAM, usually by adding GPUs.

Why VRAM is the key number

A language model's weights must fit in fast memory to answer quickly. More parameters generally mean better reasoning but more memory. On top of the weights, the system needs room for the context of each conversation, and more simultaneous users need more of that room. That is why VRAM, not CPU speed, decides what you can run.

A practical starting point

For a small team, one workstation or server with an NVIDIA GPU of 16 GB VRAM or more can run a compact model well suited to chat, summaries and document Q&A. Apple Silicon Macs (M1 or later) use unified memory shared between CPU and GPU, which makes them a convenient option for individuals and small offices. You also need fast storage for the models and the document index — plan for at least 80 GB free.

When to scale up

Move to larger models when you need stronger reasoning, long multi-step tasks or complex analysis. Add capacity when many people use the system at the same time. Instead of one huge card, servers often combine several identical GPUs so a large model can be split across them. A sensible rule: choose the smallest model that does the job well, and size hardware to it, rather than buying the biggest box first.

Matching model size to the job

Light chat and basic Q&A work well on compact models. Department-wide document assistants and business workflows benefit from mid-size models. Complex reasoning and heavy concurrent load call for the largest models on multi-GPU servers.

How DABO AI handles this for you

The DABO AI installer detects your GPU automatically and recommends a model that fits. The DABO model family ranges from DABO-Lite (12B) and DABO-Standard (26B) through DABO-Pro (27B) to DABO-Max (122B) and DABO-Ultra (284B and 320B); machines with several identical GPUs are combined automatically to run larger models. Models are not tied to plans, so even the Free plan can use whatever your hardware supports. It installs with one file on Windows (including Windows Server), Ubuntu, Red Hat or macOS, needs Internet only during installation, and runs offline afterwards. If you prefer not to buy hardware, DABO also offers managed deployments: buy outright (CAPEX), rent monthly (OPEX) or extend what you already have (Hybrid).

Frequently asked questions

What is the minimum to try private AI? An NVIDIA GPU with 16 GB VRAM or a Mac with Apple M1 or later, plus about 80 GB of free disk space.

Can I use an existing server? Yes. DABO AI installs on servers and workstations you already own if they have a suitable GPU.

Do I need a GPU at all? For a responsive experience, yes — the GPU is what makes local models fast.