Skip to content

Local AI Models

Every model Atlas uses runs on your own machine. There is no cloud model behind the app — when you chat, dictate, or hear a reply, it’s a model on your computer doing the work. This page covers which models are available, what downloads on first run, and how to manage them.

The language model is the one you chat with and the one that drives the agent tools. Atlas ships with a few options so you can match the model to your hardware. Switch between them in Settings → AI.

ModelSize (download)ContextNotes
Mistral Nemo 12B~7.5 GBup to ~32KThe default on Windows and Linux. The most capable option; needs the most memory.
Qwen3 4B~2.5 GBstandardThe lightweight choice — runs on 8 GB VRAM or a 16 GB Mac.
Qwen3 8B~5 GBstandardA middle ground between the 4B and the 12B.
Qwen3 14B~9 GBstandardLarger than the default; for machines with memory to spare.
Ministral 3 14B~8.2 GB256KA very large context window — useful for working across a lot of text at once.

On macOS the same models run as 4-bit MLX builds, which are a little smaller — the default Mistral Nemo 12B is about 6.9 GB as an MLX build, for example. Atlas downloads the right build for your platform automatically.

See Hardware Requirements for how much VRAM or unified memory each model needs. If you’re unsure, start with Qwen3 4B and move up once you know your machine handles it.

Alongside the language model, Atlas runs three smaller models. These download as part of the essential set and run on-device like everything else:

ModelPurposeSize
Whisper BaseSpeech-to-text (dictation, voice notes) — English~290 MB
KokoroText-to-speech (spoken replies)~350 MB
BGE-smallSearch embeddings (semantic search over your vault)bundled with the installer

A larger embedding model, BGE-base (~436 MB), is available if you want slightly more accurate semantic search — you can switch to it in Settings → Search & RAG. Changing the embedding model rebuilds your vault’s search index, which takes a few minutes for large vaults.

The installer is small; the models come down the first time you launch Atlas. The essential set is about 3 GB:

  • the smallest language model (Qwen3 4B),
  • Whisper Base for speech-to-text,
  • Kokoro for text-to-speech,
  • and the pronunciation dictionary Kokoro uses.

If you choose the default Mistral Nemo 12B, that brings the total to about 8 GB. The downloads are resumable, so a dropped connection picks up where it left off. You can start writing and organizing notes right away — the AI features become available once the download finishes.

Models are stored in a per-OS application-data folder, separate from your vault:

PlatformModel location
Windows%APPDATA%\Atlas-Local\models\
macOS~/Library/Application Support/Atlas-Local/models/
Linux~/.config/Atlas-Local/models/

Your vault stays untouched — it’s just your markdown notes and Atlas’s atlas/ subfolder. Keeping models out of the vault means you can back up or move your notes without dragging several gigabytes of model files along with them.

Pick a different language model any time in Settings → AI. When you switch:

  • If the new model isn’t downloaded yet, Atlas fetches it before loading.
  • The previous model’s files stay on disk, so switching back is instant — no re-download.
  • The model is loaded on demand and unloaded again after about ten minutes idle, so only the model you’re actually using occupies memory.

If a download was interrupted or a model file is corrupted, the fix is to download it again — restart the download from Settings → AI, and Atlas replaces the affected files. Because downloads are resumable and models are validated, a fresh attempt normally clears up a stuck or broken model. Your vault and license activation are never involved in this.

To free up space, keep only the language model you use and remove the others. The models you’re not running just sit on disk in the models folder listed above; deleting the files for a model you’ve switched away from reclaims its space, and Atlas will download it again if you ever switch back. The essential voice and search models are small, so the language models are where the space goes.