recipes
Install, list, and delete model recipes — curated presets the site catalogues.
ownllm recipes list
ownllm recipes install <preset>
ownllm recipes delete <id>A recipe is a curated bundle of:
- A model tag (
qwen2.5-coder:32b) - A runtime profile (
autoorapple-mlx) - Default options (
keep_alive,thinking,num_parallel) - Hardware requirements (min VRAM, RAM, recommended use case)
- A licence flag
The site maintains the catalogue at /api/v1/recipes. The CLI fetches
it on demand — no local copy to keep up to date.
list
ownllm recipes listPrints every recipe the catalogue currently exposes, filtered to the ones your hardware can run (based on the heartbeat-reported specs):
PRESET MODEL PROFILE STATUS
generalist-light llama3.2:3b auto installed
code-mid qwen2.5-coder:14b auto available
code-heavy qwen2.5-coder:32b auto installed
apple-mlx-coder qwen3.5:35b-a3b-coding apple-mlx available
vision-mid qwen2-vl:7b auto unsupported (RAM)install
ownllm recipes install code-heavyResolves the recipe → calls the equivalent of
ownllm models install <model>, applies the recipe's options, and
records the link locally so subsequent models list calls show the
recipe id.
A recipe can ship default options (keep_alive: "30m",
thinking: "low") so admins don't have to manually configure each
model after install.
delete
ownllm recipes delete code-heavyCalls ollama rm for the underlying model and removes the recipe
metadata. The model is gone after this call returns.
Why recipes?
Recipes solve two problems:
- Curation. A new tenant doesn't have to know which Ollama tag
to pull or what
keep_aliveto set. The site recommends the right model for the host's hardware. - Profiles. On Apple Silicon, MLX-accelerated recipes show up
when the agent reports
accelerators: ["metal", "mlx"]. On x86_64 + NVIDIA, those recipes are filtered out and CUDA-friendly recipes are recommended instead.
See models if you want to install something
that isn't in the catalogue.