Model Registry

The Model Registry is a typed, grouped index of every model shipped with React Native ExecuTorch. It removes the need to memorize per-model constant names: pick a capability group, pick a model, and optionally opt out of quantization or override the backend.

import { models } from 'react-native-executorch';

// Default (quantized when available, platform-default backend).
const llm = useLLM({ model: models.llm.llama3_2_3b() });

// Non-quantized variant.
const llmBase = useLLM({ model: models.llm.llama3_2_3b({ quant: false }) });

Each leaf is a function. Call it (optionally with { quant, backend }) to get the resolved model config. Accessors are not readable as values — this avoids a class of React state pitfalls (useState lazy-init, useMemo/useCallback deps), and pickers can fall back to plain === reference equality on the returned config.

Shape

models is grouped by capability. Each leaf is a callable accessor.

Group	Examples
llm	llama3_2_3b, qwen3_4b, smollm2_1_1_7b, phi_4_mini_4b, bielik_v3_0_1_5b, lfm2_5_1_2b_instruct, lfm2_5_vl_1_6b, lfm2_5_vl_450m, …
classification	efficientnet_v2_s
privacy_filter	openai, nemotron
object_detection	ssdlite_320_mobilenet_v3_large, yolo26n … yolo26x, rf_detr_nano
pose_estimation	yolo26n
semantic_segmentation	deeplab_v3_resnet50, lraspp_mobilenet_v3_large, fcn_resnet101, selfie_segmentation, …
instance_segmentation	yolo26n … yolo26x, rf_detr_nano, fastsam_s, fastsam_x
style_transfer	candy, mosaic, rain_princess, udnie
speech_to_text	whisper_tiny_en, whisper_base, whisper_small_en, …
text_to_speech	nested by model family + language: kokoro.en_us.{heart, river, sarah, adam, …}, kokoro.en_gb.{emma, daniel}, kokoro.fr.siwis, kokoro.es.{dora, alex}, kokoro.it.{sara, nicola}, kokoro.pt.{dora, santa}, kokoro.hi.{alpha, omega, psi}, kokoro.pl.mateusz, kokoro.de.anna
text_embedding	all_minilm_l6_v2, all_mpnet_base_v2, clip_vit_base_patch32_text, …
image_embedding	clip_vit_base_patch32_image
image_generation	bk_sdm_tiny_vpred_256, bk_sdm_tiny_vpred_512
vad	fsmn_vad
ocr	nested by detector: craft({ language: 'en' }) — see §OCR below

Options

type ModelOpts<B extends Backend = Backend> = {
  quant?: boolean; // Pick the non-quantized variant when false. Defaults to the quantized variant when one is published.
  backend?: B; // Explicit backend; the set of allowed values is per-model.
};

• quant defaults to the quantized variant for models that publish one (e.g. llama3_2_3b → SpinQuant, efficientnet_v2_s → int8). Pass { quant: false } to get the full-precision variant. For models with a single variant, quant is accepted but has no effect.
• backend selects an explicit backend. The accessor's call signature is typed to exactly the backends the model ships with, so requesting one the model doesn't publish is a compile-time error (e.g. models.llm.llama3_2_3b({ backend: 'coreml' }) does not type-check — Llama 3.2 is xnnpack-only). When backend is omitted, the platform-default applies: CoreML on iOS and XNNPACK on Android whenever the model ships that backend. Models that ship only one backend (e.g. xnnpack-only LLMs) use that one on both platforms.

Usage patterns

Default model

import { models } from 'react-native-executorch';

const llm = useLLM({ model: models.llm.llama3_2_3b() });
const classifier = useClassification({
  model: models.classification.efficientnet_v2_s(),
});
const stt = useSpeechToText({
  model: models.speech_to_text.whisper_tiny_en(),
});

Non-quantized variant

const llm = useLLM({ model: models.llm.qwen3_4b({ quant: false }) });
const classifier = useClassification({
  model: models.classification.efficientnet_v2_s({ quant: false }),
});

Explicit backend

// Force XNNPACK on iOS (overrides the CoreML default).
const detector = useObjectDetection({
  model: models.object_detection.rf_detr_nano({ backend: 'xnnpack' }),
});

// Combine `quant` and `backend`.
const styled = useStyleTransfer({
  model: models.style_transfer.candy({ backend: 'coreml', quant: false }),
});

Text-to-speech

text_to_speech is grouped by model family then by language code (kokoro.en_us, kokoro.en_gb, kokoro.fr, kokoro.es, kokoro.it, kokoro.pt, kokoro.hi, kokoro.pl, kokoro.de). Each leaf returns a complete Kokoro preset bundling the model, voice, and phonemizer — pass the whole result to useTextToSpeech. The kokoro level reserves room for a future TTS family without forcing a breaking rename.

import { models, useTextToSpeech } from 'react-native-executorch';

const tts = useTextToSpeech(models.text_to_speech.kokoro.en_us.heart());
// Other languages:
//   models.text_to_speech.kokoro.en_gb.emma()
//   models.text_to_speech.kokoro.fr.siwis()
//   models.text_to_speech.kokoro.pl.mateusz()

OCR

OCR is nested by detector family and parameterized by language. craft is the current pipeline (CRAFT detector + per-alphabet CRNN recognizer); future detectors land as siblings so the call site stays stable.

const ocr = useOcr({
  model: models.ocr.craft({ language: 'en' }),
});

The language parameter is type-narrowed to supported tokens ('en', 'es', 'ja', 'ko', …). See the OCR docs for the full list.

Direct imports still work

Every model is also exported as a top-level constant. Either style is supported:

import { LFM2_5_1_2B_INSTRUCT, models } from 'react-native-executorch';

useLLM({ model: LFM2_5_1_2B_INSTRUCT });
useLLM({ model: models.llm.lfm2_5_1_2b_instruct() });

Migration from the previous registry

Earlier releases exposed MODEL_REGISTRY as a flat value-only registry (MODEL_REGISTRY.ALL_MODELS.<NAME>) and shipped each precision/quantization variant as its own top-level constant. That has been replaced with a function-only, capability-grouped, lowercase models export that takes { quant, backend } opts.

// Before — flat value-only registry, one constant per variant.
useLLM({ model: MODEL_REGISTRY.ALL_MODELS.LLAMA3_2_3B });
useLLM({ model: MODEL_REGISTRY.ALL_MODELS.LLAMA3_2_3B_SPINQUANT });

// After — callable accessor; pick the variant with opts.
useLLM({ model: models.llm.llama3_2_3b() }); // quantized by default
useLLM({ model: models.llm.llama3_2_3b({ quant: false }) }); // full precision

Individual constant imports (LLAMA3_2_3B, WHISPER_TINY_EN, …) are unchanged.