Models API (Ollama Compatible)¶

Discover, describe and manage Amazon Bedrock models through the Ollama model endpoints. Served under /ollama by default; the examples below use $BASE, which includes that prefix.

At a glance¶

• Drop-in Ollama compatibility — An Ollama client's usual discovery flow — list, show, check what's resident, check the version — works by changing the base URL.
• Canonical model names — /api/tags publishes the exact names to send back as model on every other Ollama endpoint.
• Honest capability hints — capabilities reports what the catalogue actually knows about a model, never a guess dressed up as a fact.
• Private AWS backend — Backed entirely by Amazon Bedrock models in your own AWS account — no traffic to third-party endpoints.
• Differs from the Ollama API: the four store-writing verbs answer 403, /api/ps is always empty, model_info is always {}, and GET / does not answer Ollama is running — see Limits and behaviour to know.

export BASE="https://your-host/ollama"  # <scheme>://<host> + OLLAMA_ROUTES_PREFIX

curl "$BASE/api/tags" -H "Authorization: Bearer $API_KEY"

Endpoints¶

GET /api/tags¶

Lists the models this server can serve through the Ollama endpoints — a model without a chat or embedding route reachable from this dialect is not listed.

curl "$BASE/api/tags" -H "Authorization: Bearer $API_KEY"

{
  "models": [
    {
      "name": "amazon.nova-micro-v1:0",
      "model": "amazon.nova-micro-v1:0",
      "modified_at": "2024-12-03T00:00:00+00:00",
      "size": 0,
      "digest": "3f1a9c...e2b7",
      "details": {
        "parent_model": "",
        "format": "",
        "family": "Amazon",
        "families": ["Amazon"],
        "parameter_size": "",
        "quantization_level": ""
      }
    }
  ]
}

Each entry names the model by its canonical identifier — the name to send back as model on every other Ollama endpoint. size is always 0, and digest is a stable, synthetic identifier derived from the model name, 64 characters like an Ollama server's own: usable as a cache key, but explicitly not a hash of any model content, since no model file exists on this server to hash. details.format, parameter_size, quantization_level and parent_model are empty because they describe a model file this server does not have; details.family and details.families carry the model's provider instead. modified_at is the model's publication date, falling back to the Unix epoch when that date is unknown.

HEAD /api/tags answers 200 as a liveness probe, matching what an Ollama server itself answers. Like every other route it requires the API key, so it answers 401 without one; a load-balancer health check has no key to send and should target the gateway's own /health endpoint, which needs no authentication.

POST /api/show¶

Describes one model. The model is named in the request body, not the URL — POST with {"model": "..."}; name is accepted as a legacy alias of model.

curl -X POST "$BASE/api/show" \
  -H "Authorization: Bearer $API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "amazon.nova-micro-v1:0"}'

{
  "details": {
    "parent_model": "",
    "format": "",
    "family": "Amazon",
    "families": ["Amazon"],
    "parameter_size": "",
    "quantization_level": ""
  },
  "model_info": {},
  "capabilities": ["completion", "tools"],
  "modified_at": "2024-12-03T00:00:00+00:00"
}

license, modelfile, template, parameters and system are omitted entirely, and model_info is always an empty object: each describes a local model file, or GGUF metadata read out of one, and Amazon Bedrock publishes no such file or header. Rather than invent an architecture, a parameter count and a context length, this server answers the empty object Ollama uses when it knows none of them.

Capabilities¶

capabilities can report completion, tools, embedding, vision and audio, derived from the routes and modalities the catalogue already publishes for the model. It is a best-effort hint, not a contract — the backend remains the authority, and a model missing a capability here is still worth trying.

Two capabilities are never advertised:

• thinking — there is no per-model source recording whether a model reasons, so it is never claimed. think can still be sent to any model regardless: one that does not reason simply returns no thinking text.
• insert — fill-in-the-middle completion is not available on this server, on any model.

GET /api/ps¶

Lists the models currently resident in memory.

curl "$BASE/api/ps" -H "Authorization: Bearer $API_KEY"

{"models": []}

Always empty. This is the truth, not a stub: models are served on demand, so nothing is ever loaded before a request or left resident after one.

GET /api/version¶

Reports the Ollama API version this server is compatible with.

curl "$BASE/api/version" -H "Authorization: Bearer $API_KEY"

{"version": "0.33.1"}

This is a compatibility declaration, not this server's own version: an Ollama client uses it to decide which features of the Ollama API it may send. HEAD /api/version also answers 200, for clients that probe it as a liveness check — and, like every other route, it requires the API key.

Model Management¶

Every model this server offers is already available and none of them is stored locally, which leaves one verb whose post-condition can be met, and four that would have to change a model store this server does not have.

POST /api/pull¶

Answers success immediately for any model /api/tags lists — nothing is transferred, because the model is already usable — and 404 for a model this server does not offer. insecure is accepted and ignored.

curl -X POST "$BASE/api/pull" \
  -H "Authorization: Bearer $API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "amazon.nova-micro-v1:0", "stream": false}'

{"status": "success"}

By default (stream unset, or true) the same status is streamed as a single newline-delimited JSON object: {"status":"success"}. Set "stream": false to receive it as one JSON object instead, as in the example above.

POST /api/create, POST /api/copy, POST /api/push, DELETE /api/delete¶

All four always answer 403: the request is well-formed, the server simply will not perform it.

curl -X POST "$BASE/api/create" \
  -H "Authorization: Bearer $API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "my-custom-model"}'

{"error": "This server does not store models: the models it offers are hosted and already available, so they cannot be created, copied, published or deleted. Call the model list endpoint to see what is available."}

Limits and behaviour to know¶

• POST /api/create, POST /api/copy, POST /api/push and DELETE /api/delete always answer 403: this server does not store models, so these operations have no state to change, and answering 200 would tell the caller something changed when nothing did.
• GET / does not answer Ollama is running — see the note under GET /api/version above.
• capabilities is a best-effort hint derived from the catalogue, never a contract; thinking and insert are never advertised, for the reasons given under Capabilities.
• model_info is always an empty object, and details.parameter_size and details.quantization_level always empty strings: Amazon Bedrock publishes no GGUF header for the models it serves, and a plausible-looking invention is worse than the empty value Ollama itself uses for an unknown one.
• digest is the full 64-character identifier an Ollama server returns, derived from the model name rather than from any content, so it is a stable cache key and never a content hash.
• HEAD /api/tags and HEAD /api/version require the API key like every other route, so they answer 401 to an unauthenticated probe. Point a load balancer at /health instead.

Request headers¶

A local Ollama server needs no key; this one does, on every route including the HEAD probes.

Try it¶

# List what this server serves
curl "$BASE/api/tags" -H "Authorization: Bearer $API_KEY"

# Describe one model
curl -X POST "$BASE/api/show" \
  -H "Authorization: Bearer $API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "amazon.nova-micro-v1:0"}'

# Check the compatibility version
curl "$BASE/api/version" -H "Authorization: Bearer $API_KEY"

Compared with Ollama Cloud¶

Both servers answer the Ollama model endpoints for models they host rather than store, so most fields agree: format, parameter_size, quantization_level and parent_model are empty strings on both, size is 0, and license, modelfile, template, parameters and system are omitted from /api/show. Where the two differ:

Next steps¶

Next: Chat API · Generate API · Embed API · Search Models API