Ollama

Ollama can pull from any registry that speaks its dialect. The hub does, so every GGUF quantisation in the index is one ollama pull away, and Ollama checks the SHA-256 of every layer before the model is usable.

ollama pull gethologram.ai/ornith-ai/Ornith-1.5-9B-GGUF:Q4_K_M
ollama run  gethologram.ai/ornith-ai/Ornith-1.5-9B-GGUF:Q4_K_M

The name

gethologram.ai/<owner>/<name>:<quant>

<quant> is the quantisation tag of a single-file GGUF in the repository: Q4_K_M, IQ2_M, Q8_0 and so on. latest is also a tag. The tags a model has:

curl -s https://gethologram.ai/v2/ornith-ai/ornith-1.5-9b-gguf/tags/list

{ "name": "ornith-ai/Ornith-1.5-9B-GGUF", "tags": ["latest", "BF16", "Q4_K_M", "Q5_K_M", "Q6_K", "Q8_0"] }

Owner and name are lowercase on this route; OCI references are case-sensitive and lowercase-only. A model with no GGUF file has no Ollama tags and refuses the pull with a sentence saying so.

What happens

1. Ollama asks for the manifest with the Docker manifest media type. The hub answers a GGUF model manifest whose layers are the model file, its chat template and its parameters.
2. Small layers come from the hub directly.
3. The weight layer is a 307 to a source that is up right now. Ollama follows it and fetches in parallel Range parts. HEAD on the blob answers 200 with the size directly, because newer Ollama refuses a cross-host redirect on HEAD.
4. Ollama prints verifying sha256 digest. The digest it checks is the file's SHA-256 from the index.

The same three requests, by hand:

curl -s https://gethologram.ai/v2/ornith-ai/ornith-1.5-9b-gguf/manifests/Q4_K_M

{
  "schemaVersion": 2,
  "mediaType": "application/vnd.docker.distribution.manifest.v2+json",
  "config": { "digest": "sha256:84736973…", "mediaType": "application/vnd.docker.container.image.v1+json", "size": 627 },
  "layers": [
    { "digest": "sha256:70c112196e0b7023803c9762752e46d29e612a92c83f995bc3ba1ceb07e8fab6", "mediaType": "application/vnd.ollama.image.model", "size": 5780090816 },
    { "digest": "sha256:940b926e…", "mediaType": "application/vnd.ollama.image.template", "size": 201 },
    …
  ]
}

curl -sI https://gethologram.ai/v2/ornith-ai/ornith-1.5-9b-gguf/blobs/sha256:70c112196e0b7023803c9762752e46d29e612a92c83f995bc3ba1ceb07e8fab6

HTTP/1.1 200 OK
Content-Length: 5780090816
Docker-Content-Digest: sha256:70c112196e0b7023803c9762752e46d29e612a92c83f995bc3ba1ceb07e8fab6

A GET on that blob answers 307 to huggingface.co/ornith-ai/Ornith-1.5-9B-GGUF/resolve/abdd624b…/Ornith-1.5-9B-Q4_K_M.gguf, with X-Hub-Source: huggingface.co.

Ollama 0.34.2 pulled a 969 MB model this way in 67 s. With huggingface.co unreachable from the client, the same pull completed in 83 s from ModelScope, with the manifest and template served from the hub. That failover is the point of the hub, and it costs you nothing to configure.

Other engines that read the same routes

ollama run hf.co/… style names in LM Studio and Jan are hardcoded to huggingface.co and cannot be pointed here. Anything that pulls from a named OCI registry can: see OCI artifacts.