Models

Model discovery, search, listing, and custom model-card management.

📄️ Assess local fit and estimated external inference capacity

Read-only assessment of an authorized exact artifact and requested text workload. Uses production placement, reservations, context/cache estimates and compatibility. Uses observed llama-server slots and shared context pools, otherwise serial replicas. Returns local, external_required, blocked or insufficient_evidence with reasons, assumptions and hypothetical full-model external configurations. Estimates do not prove load readiness or performance. Creates no instance, reservation, provider offer or paid capacity.

📄️ Read advisory model capacity requirements

Resolve an exact known model using model-list installed/store/custom precedence and return its content binding, storage and whole-model memory estimate at the requested context. Missing text KV geometry returns a null memory estimate. No discovery, download, placement or reservation occurs. Compatibility and current identity must be rechecked on the joined node before execution.

📄️ List known models

Return the effective model catalog: complete installed cards, the current signed registry snapshot, and final operator-owned custom overrides. Entries expose active installed identity, current registry identity, update availability, verification state, and warn-only advisories.

📄️ List known models

OpenAI-style listing of Skulk's effective model catalog rather than only running instances. Entries distinguish the active installed generation from current signed registry truth and report updates, verification, advisories, and repository-code authorization provenance.

📄️ List legacy model-code approvals

Deprecated compatibility endpoint for approvals stored before model publication and explicit addition became the repository-code authorization boundaries. Returned identities are inert.

📄️ Approve model repository code (deprecated)

Deprecated compatibility endpoint. Valid signed registry cards are authorized by publication and explicitly added models are authorized by the add action, so current cards reject redundant approval with HTTP 409.

📄️ Remove a legacy model repository-code approval

Removes an inert approval retained from an older deployment. This does not revoke a published or explicitly added model.

📄️ Fetch and add a custom model card

Add a custom model card to Skulk's model catalog so it becomes searchable and launchable. An optional gguf_file selects one exact quant from a multi-quant GGUF repository. Mutable main is resolved once to an immutable commit, and the explicit add action authorizes repository code selected by that card. Success waits until the exact mutation is ordered and visible in the responding API's catalog. The mutation requires a direct loopback or trusted-fabric (private LAN / CGNAT) connection without proxy-forwarding headers, or an authenticated operator-gateway credential; browser requests must present an Origin on those trust classes or naming one of this node's own hostnames.

📄️ Add one exact unsigned custom model card

Persist a complete exact model card supplied by an authenticated operator workflow without fetching or regenerating Hub metadata. Skulk forces custom-card semantics and removes registry trust claims; the explicit immutable-card add authorizes its selected repository code without a second approval.

📄️ Delete a custom model card

Remove a previously added custom model card from Skulk's local catalog.

📄️ Search Hugging Face for models

Search for models to add or launch. Pass mlx_only=true to restrict results to mlx-community; omit or pass false to search all of Hugging Face. Exact GGUF filename queries also inspect bounded candidate repository manifests and return the matched repo-relative file path.

📄️ Fetch a Hugging Face model card summary

Download a repository's model card README and return its first prose paragraphs with markup stripped, for the dashboard's model discovery popovers. The summary is empty when the card has no usable prose.

📄️ List a GGUF repository's quantizations

Enumerate the downloadable quantizations of a Hugging Face GGUF repository: one option per quant shard group with its loadable first shard, human label, exact total bytes, and shard count, smallest first. Companion artifacts such as speculative drafters and imatrix files are excluded.