Local models¶
The native preview contains trust, installation, supervision, and decision layers for small GGUF models. The signed catalog is an empty bootstrap, so no public artifact is currently offered.
No catalog model is reference-device-qualified.
Profile before selection¶
The native command profiles operating system, architecture, CPU flags, memory, and free cache storage without downloading or executing a model. Catalog compatibility is filtered against this profile.
Inspect the catalog and cache¶
Catalog entries pin upstream revision, HTTPS URL, byte size, SHA-256 digest, GGUF format, runtime variants, memory floor, and license metadata.
Install a future qualified artifact¶
Downloads stream into a partial file, enforce declared size, verify digest, flush, and promote atomically. Cancellation, oversize response, or digest mismatch leaves no promoted model.
A gated artifact requires exact license acceptance:
JSON mode never prompts and returns license_required until the expected value is supplied.
Runtime isolation design¶
The preview selects a llama.cpp binary compatible with the machine CPU flags, verifies its
manifest, binds it to 127.0.0.1, gives it a random bearer token, checks authenticated health,
and owns the child process. An 8 GB profile defaults to one resident model and does not evict an
active lease.
Local inference is constrained to one JSON action decision with temperature zero, generated grammar, small output ceiling, and post-response schema validation. Undeclared actions and Markdown-wrapped results are rejected.
Fixture-qualified behavior is not clean-machine or reference-device evidence. Check the project qualification page before treating any future artifact as supported.