Backend covarage

This task tracks backend coverage across the models wired in the registry.

Regenerated from packages/react-native-executorch/src/models.ts on main, so it reflects what the library actually resolves rather than what exists on Hugging Face. 104 models across 15 task groups.

Columns

• CPU (XNNPACK) runs on both platforms; it is the portable fallback and every model has it.
• Android GPU (Vulkan) is Android only.
• iOS (Core ML / MLX) is Apple only. Core ML reaches ANE/GPU/CPU; MLX targets Apple Silicon GPU.

Cells list the published precision variants, so 8da4w, fp16 means both are wired. — means the model has no variant for that column.

LLMs

Model	CPU (XNNPACK)	Android GPU (Vulkan)	iOS (Core ML / MLX)
LFM2_5_1_2B	8da4w, fp16	—	MLX int4
LFM2_5_350M	8da4w, fp16	—	MLX int4
LFM2_5_VL_450M	8da4w	8da4w	MLX int4
LFM2_5_VL_1_6B	8da4w	8da4w	MLX int4, int8
BIELIK_V3_1_5B	8da4w, fp16	—	—
LLAMA3_2_1B	spinquant, bf16	—	MLX int4
LLAMA3_2_3B	spinquant, bf16	—	MLX int4
SMOLLM2_135M	8da8w	—	MLX int8
SMOLLM2_360M	8da8w	—	MLX int8
SMOLLM2_1_7B	8da8w	—	MLX int8
HAMMER2_1_0_5B	8da4w, bf16	—	MLX int4
HAMMER2_1_1_5B	8da4w, bf16	—	MLX int4
HAMMER2_1_3B	8da4w, bf16	—	MLX int4
PHI4_MINI	8da4w, bf16	—	MLX int4
QWEN2_5_0_5B	8da4w, bf16	—	MLX int4
QWEN2_5_1_5B	8da4w, bf16	—	MLX int4
QWEN2_5_3B	8da4w, bf16	—	MLX int4
QWEN3_0_6B	8da4w, bf16	—	MLX int4
QWEN3_1_7B	8da4w, bf16	—	MLX int4
QWEN3_4B	8da4w, bf16	—	MLX int4
GEMMA4_E2B	8da4w	8da4w	MLX int4

Text embeddings

Model	CPU (XNNPACK)	Android GPU (Vulkan)	iOS (Core ML / MLX)
ALL_MINILM_L6_V2	fp32	fp16	Core ML fp16
ALL_MPNET_BASE_V2	fp32	fp16, int8	—
MULTI_QA_MINILM_L6_COS_V1	fp32	fp16	Core ML fp16
MULTI_QA_MPNET_BASE_DOT_V1	fp32	fp16, int8	—
PARAPHRASE_MULTILINGUAL_MINILM_L12_V2	8da4w, fp32	fp16	Core ML fp16
DISTILUSE_BASE_MULTILINGUAL_CASED_V2	8da4w, fp32	fp16	Core ML fp16 · MLX int8
CLIP_VIT_BASE_PATCH32_TEXT	fp32	fp16	Core ML fp16
LFM2_5_EMBEDDING_350M	8da4w	—	MLX int4

Image embeddings

Model	CPU (XNNPACK)	Android GPU (Vulkan)	iOS (Core ML / MLX)
CLIP_VIT_BASE_PATCH32	fp32	fp16	Core ML fp16 · MLX int8

Speech to text

Model	CPU (XNNPACK)	Android GPU (Vulkan)	iOS (Core ML / MLX)
WHISPER.TINY	fp32	fp16, int8	Core ML fp16 · MLX bf16, int8
WHISPER.BASE	fp32	fp16, int8	Core ML fp16 · MLX bf16, int8
WHISPER.SMALL	fp32	fp16, int8	Core ML fp16 · MLX int8
WHISPER.EN.TINY	fp32, int8	fp16, int8	Core ML fp16 · MLX bf16, int8
WHISPER.EN.BASE	int8, fp32	fp16, int8	Core ML fp16 · MLX bf16, int8
WHISPER.EN.SMALL	int8, fp32	fp16, int8	Core ML fp16 · MLX int8

Text to speech

Model	CPU (XNNPACK)	Android GPU (Vulkan)	iOS (Core ML / MLX)
SUPERTONIC	fp32	fp16	MLX fp32
KOKORO.EN_US	fp32	—	Core ML fp32
KOKORO.EN_GB	fp32	—	Core ML fp32
KOKORO.ES	fp32	—	Core ML fp32
KOKORO.FR	fp32	—	Core ML fp32
KOKORO.IT	fp32	—	Core ML fp32
KOKORO.PT	fp32	—	Core ML fp32
KOKORO.HI	fp32	—	Core ML fp32
KOKORO.PL	fp32	—	Core ML fp32
KOKORO.DE	fp32	—	Core ML fp32

Voice activity detection

Model	CPU (XNNPACK)	Android GPU (Vulkan)	iOS (Core ML / MLX)
FSMN_VAD	fp32	—	—

Image classification

Model	CPU (XNNPACK)	Android GPU (Vulkan)	iOS (Core ML / MLX)
EFFICIENTNET_V2_S	int8, fp32	—	Core ML fp16

Object detection

Model	CPU (XNNPACK)	Android GPU (Vulkan)	iOS (Core ML / MLX)
SSDLITE320_MOBILENET_V3_LARGE	fp32	—	Core ML fp16
RFDETR_NANO	fp32	—	Core ML fp16
YOLO26.NANO.SIZE_384	fp32	—	Core ML fp16
YOLO26.NANO.SIZE_512	fp32	—	Core ML fp16
YOLO26.NANO.SIZE_640	fp32	—	Core ML fp16
YOLO26.SMALL.SIZE_384	fp32	—	Core ML fp16
YOLO26.SMALL.SIZE_512	fp32	—	Core ML fp16
YOLO26.SMALL.SIZE_640	fp32	—	Core ML fp16
YOLO26.MEDIUM.SIZE_384	fp32	—	Core ML fp16
YOLO26.MEDIUM.SIZE_512	fp32	—	Core ML fp16
YOLO26.MEDIUM.SIZE_640	fp32	—	Core ML fp16
YOLO26.LARGE.SIZE_384	fp32	—	Core ML fp16
YOLO26.LARGE.SIZE_512	fp32	—	Core ML fp16
YOLO26.LARGE.SIZE_640	fp32	—	Core ML fp16
YOLO26.XLARGE.SIZE_384	fp32	—	Core ML fp16
YOLO26.XLARGE.SIZE_512	fp32	—	Core ML fp16
YOLO26.XLARGE.SIZE_640	fp32	—	Core ML fp16

Keypoint detection

Model	CPU (XNNPACK)	Android GPU (Vulkan)	iOS (Core ML / MLX)
BLAZEFACE	fp32	—	—
YOLO26_POSE.SIZE_384	fp32	—	Core ML fp16
YOLO26_POSE.SIZE_512	fp32	—	Core ML fp16
YOLO26_POSE.SIZE_640	fp32	—	Core ML fp16
RFDETR_KEYPOINT	fp32	—	Core ML fp16

Instance segmentation

Model	CPU (XNNPACK)	Android GPU (Vulkan)	iOS (Core ML / MLX)
FASTSAM.S	fp32	—	Core ML fp16
FASTSAM.X	fp32	—	Core ML fp16
RFDETR_NANO	fp32	—	Core ML fp16
YOLO26.NANO.SIZE_384	fp32	—	Core ML fp16
YOLO26.NANO.SIZE_512	fp32	—	Core ML fp16
YOLO26.NANO.SIZE_640	fp32	—	Core ML fp16
YOLO26.SMALL.SIZE_384	fp32	—	Core ML fp16
YOLO26.SMALL.SIZE_512	fp32	—	Core ML fp16
YOLO26.SMALL.SIZE_640	fp32	—	Core ML fp16
YOLO26.MEDIUM.SIZE_384	fp32	—	Core ML fp16
YOLO26.MEDIUM.SIZE_512	fp32	—	Core ML fp16
YOLO26.MEDIUM.SIZE_640	fp32	—	Core ML fp16
YOLO26.LARGE.SIZE_384	fp32	—	Core ML fp16
YOLO26.LARGE.SIZE_512	fp32	—	Core ML fp16
YOLO26.LARGE.SIZE_640	fp32	—	Core ML fp16
YOLO26.XLARGE.SIZE_384	fp32	—	Core ML fp16
YOLO26.XLARGE.SIZE_512	fp32	—	Core ML fp16
YOLO26.XLARGE.SIZE_640	fp32	—	Core ML fp16

Semantic segmentation

Model	CPU (XNNPACK)	Android GPU (Vulkan)	iOS (Core ML / MLX)
SELFIE_SEGMENTATION	fp32	—	Core ML fp16
SELFIE_SEGMENTATION_LANDSCAPE	fp32	—	Core ML fp16
LRASPP_MOBILENET_V3_LARGE	int8, fp32	—	Core ML fp16
DEEPLAB_V3_RESNET50	int8, fp32	—	Core ML fp16
DEEPLAB_V3_RESNET101	int8, fp32	—	Core ML fp16
DEEPLAB_V3_MOBILENET_V3_LARGE	int8, fp32	—	Core ML fp16
FCN_RESNET50	int8, fp32	—	Core ML fp16
FCN_RESNET101	int8, fp32	—	Core ML fp16

Style transfer

Model	CPU (XNNPACK)	Android GPU (Vulkan)	iOS (Core ML / MLX)
CANDY	int8, fp32	—	Core ML fp16
MOSAIC	int8, fp32	—	Core ML fp16
RAIN_PRINCESS	int8, fp32	—	Core ML fp16
UDNIE	int8, fp32	—	Core ML fp16

OCR

Model	CPU (XNNPACK)	Android GPU (Vulkan)	iOS (Core ML / MLX)
PADDLE.PPOCRV6_SMALL	yes, fp32	yes	Core ML yes

Text to image

Model	CPU (XNNPACK)	Android GPU (Vulkan)	iOS (Core ML / MLX)
SDXS_512_DREAMSHAPER	fp32	—	Core ML fp16

Privacy filter

Model	CPU (XNNPACK)	Android GPU (Vulkan)	iOS (Core ML / MLX)
OPENAI	8da4w	—	MLX int4
NEMOTRON	8da4w	—	MLX int8

---

Notes

• Vulkan on LFM2.5 is the VL variants only (LFM2_5_VL_450M, LFM2_5_VL_1_6B); the text-only 350M and 1.2B are XNNPACK plus MLX.
• RF-DETR keypoint intentionally has no MLX variant: on an iPhone 16, Core ML fp16 measured 141.6 ms / 248 MB against MLX fp32 at 382.2 ms / 1153 MB, and the MLX delegate has no quantized convolution to close the gap.
• Core ML text embedders accept a 1-token input as of the seq=1 export fix (react-native-executorch#1164).

References

• Building Voice Agents with ExecuTorch (PyTorch blog)