Get a model onto Umbra

Umbra's catalog grows one model at a time as providers add them. Here is how a model reaches the network, whether you host it yourself or want someone else to.

Public weights only. Every model on Umbra is a public, open-weight model on Hugging Face, in GGUF form, pulled by a provider with their own Hugging Face key. There are no buyer uploads and no secret or private weights. That invariant is what keeps the network a neutral pipe.

The short version

Models are added by providers, not uploaded through the website. Hosting is an on-device job run from the Umbra CLI on an Apple Silicon Mac. If a model you want is already a public GGUF on Hugging Face, any provider can bring it online with a single command.

If you run a provider Mac

Pick any public GGUF repository on Hugging Face and host it from the CLI:

umbra host <hf-model-id>

Not sure what fits your machine or what it could earn? The host a model page ranks which GGUF models earn the most on your exact hardware, and its paste-any-link box prices any Hugging Face GGUF repo on your Mac before you commit. The model must ship .gguf files, since llama.cpp serves GGUF only; a repo with just the original safetensors or PyTorch weights cannot be hosted as-is, but community GGUF builds of popular models are common.

Behind the scenes the onboarding flow validates the repo, checks the architecture against the allowlist, pulls or converts to GGUF and records the integrity digest, then registers and approves the entry, all with your own Hugging Face key. Once added, the model appears under the models you host and serves whenever your device is attested and online.

Architecture allowlist

For safety and reliability, Umbra hosts a fixed set of model architectures: llama, qwen2, qwen3, gemma, gemma2, gemma3, mistral, mixtral, and phi3. A repo whose architecture is not on the allowlist is rejected with a clear message. This list grows over time as more architectures are validated.

If you want a model but do not host

You have two paths:

Publish it and ask. Put a public GGUF build of the model on Hugging Face, then reach out via the contact page with the repo id so a provider can pick it up.
Become a provider. The most direct route is to host it yourself on an Apple Silicon Mac. See become a provider to get started.

Pricing and approval

Pricing is platform-set: every provider of a given model earns the same per-model input and output rate, served live from GET /v1/models. Providers choose and approve which models they host and are responsible for each model's license and the laws that apply to them, as described in the Provider Agreement.

Ready to add a model? Open the host a model page to price a GGUF on your Mac, or read the quickstart to call the models already live.