# Registry List, inspect, share and delete your org's models and containers over HTTP. The [registry](https://docs.privatemind.com/registry.html) holds your org's model weights and container images. This API manages what is in it: listing, reading a version's files and README, switching private and shared, handing over ownership, and deleting. Pushing and pulling go through the registry protocol at `privatemind.com/v2/`, with the [`privatemind` command](https://docs.privatemind.com/sdk-weights.html#the-command-line) for models and Docker for containers. Models are on `https://privatemind.com/api/v1/weights` and containers on `https://privatemind.com/api/v1/images`. The two take the same calls, except that a container's versions are called tags. ## Authentication Send a personal API key in the `Authorization` header, access key id and secret joined by a colon: ```text Authorization: Bearer : ``` - **Models** also take a workspace or job key. It acts for its owner, with a member's rights: it can read and push. It can't delete or change sharing or ownership. - **Containers** take personal keys only. - `Inference only`, `Inference and containers` and management keys get `403`. A read-only key can only read. - What you see follows the [private and shared rules](https://docs.privatemind.com/registry.html#private-and-shared): another member's private model answers `403`. If the registry is not on for your org, every call answers `403` saying so. ## List ```bash curl -s https://privatemind.com/api/v1/weights \ -H "Authorization: Bearer $PRIVATEMIND_API_KEY" ``` ```json { "success": true, "body": [ { "kind": "weights", "name": "qwen3-0.6b", "path": "models/your-org/qwen3-0.6b", "owner": { "id": 12, "name": "Jane Doe" }, "visibility": "private", "description": null, "version_count": 1, "size_bytes": 1519227280, "last_pushed_at": "2026-11-17T05:32:38.708Z", "latest": { "version": "2026-11-17", "digest": "sha256:f0a9800a…", "pushed_at": "2026-11-17T05:32:38.708Z", "size_bytes": 1519227280, "platforms": ["linux/amd64", "linux/arm64"] }, "pulls": 0, "last_pulled_at": null, "card": { "task": "text-generation", "architecture": "Qwen3ForCausalLM", "model_type": "qwen3", "parameters": 751632384, "precision": "BF16", "quantization": null, "context_length": 40960, "experts": null, "format": "safetensors", "file_count": 9, "has_tokenizer": true, "has_chat_template": true, "adapter_base": null, "license": "apache-2.0", "base_model": "Qwen/Qwen3-0.6B-Base" }, "can": { "push": true, "delete": true, "set_visibility": true, "transfer": false } } ], "total": { "path": "models/your-org", "used_bytes": 1883308800, "quota_bytes": 107374182400 } } ``` `body` holds every model you can read, and `total` is your org's usage and quota. `can` says what you may do with each one. `card` describes the newest version from its own files, as [What a version shows](https://docs.privatemind.com/registry.html#what-a-version-shows) explains. It is `null` for containers. It is also `null` for a model version pushed with another registry tool that did not name each file (the `org.opencontainers.image.title` annotation). The `privatemind` command always names them. A version pushed a moment ago can come back with `card: null` once and gain it on the next call. `pulls` counts downloads across versions, and `last_pulled_at` is `null` when nobody has pulled one. ## One model or container ```bash curl -s https://privatemind.com/api/v1/weights/qwen3-0.6b \ -H "Authorization: Bearer $PRIVATEMIND_API_KEY" ``` The answer's `total` carries the same fields as a list entry, plus `versions`, newest first. Each version looks like this: ```json { "version": "2026-11-17", "digest": "sha256:f0a9800a…", "size_bytes": 1519227280, "pushed_at": "2026-11-17T05:32:38.708Z", "platforms": [ { "os": "linux", "arch": "amd64", "variant": null, "size_bytes": 1519227280 }, { "os": "linux", "arch": "arm64", "variant": null, "size_bytes": 1519227280 } ], "pulls": 0, "last_pulled_at": null, "used_by": [ { "kind": "workspace", "name": "eval-7f2k", "display_name": "Qwen eval", "owner_user_id": 12, "owner_name": "Jane Doe" } ], "used_by_hidden": 0 } ``` `used_by` lists the workspaces, jobs and model servers that use the version. Another member's private workload is counted in `used_by_hidden` and not named. ## One version `GET /api/v1/weights/{name}/versions/{version}` for a model, `GET /api/v1/images/{name}/tags/{tag}` for a container. A model version answers its `card`, its `files` (each `path` and `size_bytes`, sorted by path) and its `readme`: ```json { "success": true, "total": { "kind": "weights", "name": "qwen3-0.6b", "version": "2026-11-17", "digest": "sha256:f0a9800a…", "card": { "task": "text-generation", "architecture": "Qwen3ForCausalLM", "parameters": 751632384, "precision": "BF16" }, "files": [ { "path": "config.json", "size_bytes": 726 }, { "path": "model.safetensors", "size_bytes": 1503300328 } ], "files_truncated": false, "readme": "# Qwen3-0.6B\n\n…", "readme_truncated": false } } ``` The file list stops at 1000 files, or sooner when the paths are very long. `readme` stops at 64 KiB of `README.md`, with its front matter removed. The matching `_truncated` field says when either was cut. The README is the author's markdown, so render it as untrusted content. A container tag answers its `platforms`, its layer count and its image `config`: `entrypoint`, `cmd`, `working_dir`, `user`, `exposed_ports` and `labels`. The image's environment variables are never returned. ## Share, describe or hand over ```bash curl -s -X PATCH https://privatemind.com/api/v1/weights/qwen3-0.6b \ -H "Authorization: Bearer $PRIVATEMIND_API_KEY" \ -H "Content-Type: application/json" \ -d '{"visibility": "shared"}' ``` | Field | Who may send it | Notes | |---|---|---| | `visibility` | the owner or an org admin | `"private"` or `"shared"`. Going private answers `409` while another member's workload uses it, naming the workloads. | | `description` | the owner or an org admin | Text up to 2000 characters, or `null` to clear it. | | `owner_user_id` | an org admin | A member of your org, or `null` to make the org the owner. A private model cannot change owner while a workload the new owner does not own uses it: that answers `409`. | The answer is the updated model, as in a list entry. ## Delete ```bash curl -s -X DELETE https://privatemind.com/api/v1/weights/qwen3-0.6b/versions/2026-11-17 \ -H "Authorization: Bearer $PRIVATEMIND_API_KEY" ``` | Call | What it deletes | |---|---| | `DELETE /api/v1/weights/{name}/versions/{version}` | One model version. | | `DELETE /api/v1/weights/{name}` | A model and every version. | | `DELETE /api/v1/images/{name}/tags/{tag}` | One container tag. | | `DELETE /api/v1/images/{name}` | A container and every tag. | The owner or an org admin may delete. For a container pushed before private and shared existed (`legacy_open` is `true`), any member with compute access may delete too. If a workspace, job or model server uses the version, running or stopped, the call answers `409` with `used_by` and `used_by_hidden`, as above, and deletes nothing. Deleting a whole model or container is refused the same way while any of its versions is in use. ## Errors | Status | When | |---|---| | `400` | A field that is not valid, with the reason in `error`. | | `403` | The registry is not on for your org, the key cannot make this call, or the model or container is private to another member. | | `404` | No such model, container, version or tag. | | `409` | It is in use. `used_by` names the workloads. | | `502` | The registry did not answer. Try again. | | `503` | The in-use check could not run. Try again shortly. | Most errors answer `{"success": false, "error": "…"}`. A refused key gets `message` in place of `error`, and a workspace or job key missing a scope gets `error` as an object with `type`, `message` and `required_scope`. See [Errors](https://docs.privatemind.com/errors.html) for the rest of the platform's codes.