Run models
Skippy can serve a model from its catalog, an exact Hugging Face repository reference, or an existing local file. The source identifies the model bytes; the runtime and serving settings determine how those bytes execute.
Use a catalog name for the shortest supported path, a Hugging Face reference when you need a particular repository and quantization, and a local path for models you already manage.
Choose a model source
skippy serve --model Qwen3-0.6B-Q4_K_M
skippy serve --model unsloth/Qwen3-8B-GGUF:Q4_K_M
skippy serve --model /models/qwen3-8b-q4_k_m.ggufRemote GGUF files resolve to an immutable revision and are cached before loading. Serving the same exact reference reuses the Hugging Face cache.
Discover and manage models
Skippy uses the Mesh LLM catalog for curated names and recommendations by default:
skippy models recommendedrecommended shows curated starting points from the selected catalog.
You can provide your own catalog by creating a Hugging Face dataset repository with the same format as meshllm/catalog , then setting the repository as an override:
export SKIPPY_MODEL_CATALOG=your-org/your-catalog
skippy models recommendedCatalog caches are kept separately by repository, so changing the override does not mix entries from different catalogs.
Use search when you know a family or keyword but not the exact repository:
skippy models search qwen --limit 10Search queries Hugging Face for GGUF repositories by default. Once you have a candidate, resolve the exact artifact, quantization, revision, and expected files before downloading it:
skippy models show unsloth/Qwen3-8B-GGUF:Q4_K_M
skippy models download unsloth/Qwen3-8B-GGUF:Q4_K_Mshow is read-only. download verifies and stores the selected files in the
shared Hugging Face cache so later serving can reuse them without another
download.
Inspect and maintain the local cache separately:
skippy models installed
skippy models updates --check
skippy models cleanup --unused-since 30dinstalled reports the model references Skippy can resolve locally.
updates --check compares cached references with upstream revisions without
downloading changes. cleanup finds managed model files that have not been used
within the selected period.
Cleanup and deletion preview their targets by default. Review the selected
files before adding --yes; Skippy never deletes a remote repository.
Automatic resource planning
For a local GGUF, Skippy sizes one unified KV pool from model metadata, weight footprint, and available device memory, then chooses context and concurrency defaults. Override those decisions only when you have a measured reason:
skippy serve \
--model unsloth/Qwen3-8B-GGUF:Q4_K_M \
--ctx-size 32768 \
--generation-concurrency 2Inspect the resolved launch before model load with:
skippy serve --settings serve.toml --print-effective-configHow model resolution works
Skippy resolves a remote reference to an immutable repository revision, selects the requested artifacts, verifies the downloaded files, and records the resulting model identity. Local files are read directly but still pass model admission and memory planning before the server becomes ready.
For local GGUF models, the planner uses model metadata, weight footprint, and
available device memory to choose a context size and generation concurrency.
An explicit --ctx-size or --generation-concurrency overrides the matching
automatic choice. If the requested model or context does not fit, Skippy fails
the load instead of silently substituting another model.
Model command reference
| Task | Command |
|---|---|
| Show recommended models | skippy models recommended |
| Search Hugging Face | skippy models search QUERY |
| Resolve an exact artifact | skippy models show REFERENCE |
| Download before serving | skippy models download REFERENCE |
| List cached models | skippy models installed |
| Check cached refs for updates | skippy models updates --check |
| Preview unused-file cleanup | skippy models cleanup --unused-since 30d |
Skippy honors the standard Hugging Face cache and authentication settings.
Model deletion and cleanup are previews until --yes is supplied.