Model registry

kaos-nlp-transformers only loads models from a curated registry — each entry is license-reviewed and SHA-pinned (never a moving main revision), so a model can’t silently change underneath you. The tasks and their registered families:

Task	Registered models (examples)	Notes
Embeddings	minishlab/potion-base-8M, potion-base-32M, potion-retrieval-32M (static, [model2vec]); BAAI/bge-small-en-v1.5 (ONNX)	static models load with no download — used in the embeddings how-to
Reranking	cross-encoder rerankers	re-score query/passage pairs after retrieval
NLI	small NLI cross-encoders	entailment scoring; satisfies the classify protocol
NER	GLiNER-style zero-shot extractors	label-driven span extraction
PII	a closed-label BERT-small token classifier	~24 PII categories, faster than zero-shot NER

How loading works

• The ort (ONNX Runtime) Rust cdylib is the canonical inference backend; model2vec is the static-numpy backend for registered static embeddings. No PyTorch in the base runtime path.
• Models download on first use into a cache; pre-warm with kaos-nlp-transformers prefetch and run offline with KAOS_NLP_TRANSFORMERS_OFFLINE=1. The vendored static models need no download at all.

Bypassing the registry

You can load an unregistered model with KAOS_NLP_TRANSFORMERS_ALLOW_UNREGISTERED=true — but then license review and contract-matching are your responsibility. The registry exists precisely so the default path is safe. See the determinism contract for why pinning matters.