Hardware Requirements
Atlas runs its language model, speech-to-text, and text-to-speech on your own machine — nothing is sent to a server. That is the whole point of the local-first design, but it does mean the AI features need a capable GPU. Your notes, editing, search, and everything that isn’t the chat model will run on almost anything; the language model is what has real requirements.
The short version
Section titled “The short version”- Windows and Linux: an NVIDIA GPU with CUDA. The amount of video memory (VRAM) decides which model you can run.
- macOS: Apple Silicon (M1 or later). The amount of unified memory decides which model you can run.
- No dedicated GPU? You can still use the vault, notes, and search. The chat model won’t run — there is no CPU fallback for the language model.
The rest of this page helps you pick a model for the hardware you have.
GPU memory and which model to run
Section titled “GPU memory and which model to run”The single most important number is how much VRAM (on NVIDIA) or unified memory (on Apple Silicon) you have. Bigger models are more capable but need more memory, and a larger context window (how much conversation and note text the model can consider at once) needs more memory again.
NVIDIA (Windows / Linux)
Section titled “NVIDIA (Windows / Linux)”| Your VRAM | What to run |
|---|---|
| 24 GB or more | Mistral Nemo 12B (the default) with a large context window (~32K) |
| 16 GB | Mistral Nemo 12B at a moderate context window (~16K) |
| 10–12 GB | Mistral Nemo 12B works, but it’s tight — keep the context window smaller |
| 8 GB | Choose Qwen3 4B instead of the 12B |
As a rule of thumb, the default 12B model needs roughly 10–12 GB of VRAM at a 16–32K context window. Pushing the context window higher costs more memory — around 12 GB at 32K and around 17 GB at 64K with the 12B model. If a model fails to load or you run out of memory, drop to a smaller model or a smaller context window in Settings → AI.
Apple Silicon (macOS)
Section titled “Apple Silicon (macOS)”On a Mac, the GPU shares the same memory as the rest of the system (“unified memory”), so the number to look at is your total RAM.
| Your unified memory | What to run |
|---|---|
| 24–32 GB or more | Mistral Nemo 12B (the default) |
| 16 GB | Choose Qwen3 4B |
Apple Silicon runs the model on MLX/Metal, so any M-series Mac with enough memory works — you do not need a discrete GPU.
Memory for the other components
Section titled “Memory for the other components”The language model is by far the biggest consumer of memory, but Atlas also loads smaller models for search and voice:
- Search (embeddings): the bundled BGE-small embedding model.
- Speech-to-text: Whisper Base.
- Text-to-speech: Kokoro.
- Wake word: the OpenWakeWord detector, if you turn “Hey Jarvis” on.
Together these add up to roughly 1.5 GB of memory when voice and search are all active at once — small next to the language model. Atlas loads them on demand and unloads them again after about ten minutes of not being used, so they only occupy memory when you’re actually using them. The first request after an idle stretch is a little slower because the model has to reload.
Disk space
Section titled “Disk space”The Atlas installer itself is small. The AI models are downloaded separately the first time you launch the app:
- ~3 GB for the essential set — the smallest language model (Qwen3 4B), Whisper Base for speech-to-text, and Kokoro for text-to-speech.
- ~8 GB if you use the default Mistral Nemo 12B model instead of the 4B.
Leave room for your vault and its search index on top of that. Your vault is just plain markdown files, so it’s as large as the notes you put in it. See Local AI Models for the full model lineup and sizes.
Will my machine run this?
Section titled “Will my machine run this?”A quick way to check before you buy:
- Windows / Linux, NVIDIA card with 8 GB or more: yes. With 8 GB, run Qwen3 4B; with 10–12 GB or more, run the default 12B.
- Any Apple Silicon Mac (M1 or later) with 16 GB or more: yes. With 16 GB, run Qwen3 4B; with 24 GB or more, run the 12B.
- Windows / Linux with only integrated graphics, or an older Intel Mac: the notes, search, and vault features work, but the chat model won’t run.
If you’re on the edge — say an 8 GB card, or a 16 GB Mac — start with Qwen3 4B. It runs comfortably on lighter hardware and still handles chat, tools, and voice; you can always switch up to a bigger model later in Settings → AI.