Models

Atlas talks to the local Ollama API to manage models, then mirrors the state to the OwnLLM site through the heartbeat. Anything you do in Atlas is reflected on the admin web within ~30 seconds; anything an admin does on the web shows up on the next heartbeat.

Install a recipe

The recipes the OwnLLM team curates appear under Recommended:

1. Filter by use case (Generalist, Code, Vision, Long context).
2. Read the hardware requirements — Atlas hides recipes the host can't run.
3. Click Install. Atlas downloads the model in the background and shows a per-recipe progress bar.

A recipe may include default options (keep_alive: 30m, thinking: low). Atlas applies them after install — you don't have to configure anything.

Install a custom Ollama tag

Click + Install custom and paste any Ollama tag:

phi4:14b
qwen3:32b-a8b-thinking-q5
deepseek-r1:32b

Atlas calls ollama pull <tag> and surfaces the same progress bar. Custom installs don't get the curation guarantees (license check, hardware fit, capability annotations) that recipes do — you're on the hook for picking something the host can run.

Remove a model

In the model list, click the row → Remove. Atlas calls ollama rm and clears the local recipe link if any.

The model is gone from Ollama immediately; the next heartbeat notifies the site so the model disappears from the chat web's model selector.

Per-model options

Today, options live in the CLI: ownllm models config <id>. The Atlas dashboard shows the current values but editing them moves you to the terminal in this release.

The recognised options are:

• keep_alive — 0, 5m, 30m, -1 (pin in VRAM). Default 5m.
• thinking — default, disabled, low, medium, high. Only applies to thinking-capable models (Qwen3, DeepSeek-R1).
• num_parallel — concurrent requests this model accepts before queuing. Default: 2.

A future Atlas release will surface these as native UI controls.

Apple Silicon — MLX recipes

If Atlas is running on a Mac with Apple Silicon, the heartbeat reports accelerators: ["metal", "mlx"]. Recipes flagged runtimeProfile: "apple-mlx" then become installable; on x86_64 machines they're hidden.

The first MLX recipe is apple-mlx-coder (qwen3.5:35b-a3b-coding-nvfp4) — recommended for code on machines with 32 GB+ unified memory.

Disk pressure

Models are big: a 70B model is ~40 GB on disk. Atlas warns when free disk drops below 20 GB and refuses to install when below 10 GB. The warning appears in the dashboard's Status panel.

Where the files live

Atlas asks Ollama to download to its default location:

To clean up disk, use Remove in the Atlas UI or ollama rm <model> from a terminal — don't delete the files directly, or Ollama's index drifts.