Loaded Resources

GET /v2/resources?view=Loaded lists the resources the fleet currently holds: cached on at least one worker and usable now, with no download to wait for. Use it to keep an external catalogue in step with what the orchestrator can run. Its sibling view=Queue lists what is being loaded.

curl -G https://orchestration.civitai.com/v2/resources \
  -H "Authorization: Bearer $CIVITAI_TOKEN" \
  -d view=Loaded \
  -d take=100 \
  -d source=civitai \
  -d minSizeBytes=1073741824

Items are ResourceInfo, the same shape GET /v2/resources/{air} returns, paged with the usual next cursor.

Filter first — an unfiltered query is refused

The fleet holds hundreds of thousands of loaded resources, mostly not models but input images, sample images and other workflow artifacts. When more than 50,000 resources match on one provider, the request is refused with a 400 saying how many matched:

{N} resources match on one provider. Narrow the query with type, source or minSizeBytes.

The filters compose with AND; each repeatable one is OR'd within itself.

Parameter	Matches
type	The AIR's type segment. Repeatable.
excludeType	The AIR's type segment; wins over type. Repeatable.
source	The AIR's source segment. Repeatable.
minSizeBytes / maxSizeBytes	The resource's size in bytes, inclusive.

Filtering by type, and why checkpoint is not "every base model"

type matches the AIR's type segment literally: no alias expansion, no enum, no validation. An unrecognised value matches nothing.

The segment comes from the model's type, not the file's (a Diffusion Model file on a Checkpoint model is published as …:checkpoint:…), so base weights are spread across several segments:

• checkpoint
• diffusion_model and diffusionmodel
• unet — GGUF weights, a distinct loader slot from diffusion weights

type=checkpoint alone returns only some base models, and the segments are not aliased server-side. Pass each one you want:

?view=Loaded&type=checkpoint&type=diffusion_model&type=diffusionmodel&type=unet

Filtering by size

minSizeBytes is the blunter but more reliable cut when you want "the big models" rather than a category:

?view=Loaded&minSizeBytes=1073741824

1 GiB (1073741824) is the meaningful boundary: the orchestrator puts larger resources in its large download lane, and in practice it separates base weights from LoRAs, embeddings and VAEs.

A size of 0 means unknown, not empty

Resources whose provider never reported a byte count have size: 0, so any minSizeBytes excludes them. Filter by type or source to include them.

Filtering by source

source=civitai is the cheapest single predicate for a catalogue sync. It drops everything the orchestrator generated or fetched itself: blob and object-storage URLs, represented by synthetic AIRs with source orchestrator and type other, which are the bulk of the table.

To exclude them by type instead, the non-model artifacts are other, torchcompilecache and nodepacklayer:

?view=Loaded&excludeType=other&excludeType=torchcompilecache&excludeType=nodepacklayer

A complete list, or an error

view=Loaded never returns a partial list. If any part of the fleet is unreachable, or still re-registering workers after an orchestrator restart, the response is 503 with Retry-After. Treat anything but 200 as "no answer this time", never as "these resources were unloaded".

Paging

Pass each response's next as cursor. take defaults to 100, which is also the maximum.

Each page is a fresh snapshot, not a slice of one consistent read: if the fleet's cache changes mid-sweep, an item can be missed or repeated at a page boundary. Make your reconciliation idempotent. view=Queue pages the same way.

The same filters on view=Queue

?view=Queue&type=checkpoint narrows the download queue the same way, with two differences:

• The queue does not carry sizes, so minSizeBytes / maxSizeBytes make the orchestrator look up each queued candidate's size. Filtering by type or source is free.
• There is no too-broad refusal: the queue holds only what is being pulled right now.

What this view does not tell you

• availability is not populated — every item is available by definition. Call GET /v2/resources/{air} for one resource's detail.
• Nothing about capacity. Worker counts, lane occupancy and download bandwidth are not exposed.