Supported Models

Every model below ships pre-exported in the models registry and downloads on first use. Backends: XNNPACK (CPU, iOS and Android), Core ML and MLX (iOS), Vulkan (Android GPU); DEFAULT picks the fastest one the device can run. Any other ExecuTorch .pte can be loaded too, see Exporting Custom Models.

LLMs & VLMs

Guide · Benchmarks

Chat sessions stream tokens and support tool calling, a thinking-mode switch, Jinja chat templates, constrained sampling and image/audio inputs on multimodal models.

Model	Description	Backends	Precisions	Added in
models.llm.MINICPM5_2B	OpenBMB MiniCPM5 2B chat model with an 8K context window.	XNNPACK, MLX	8da4w, int4	v0.11.0
models.llm.LFM2_5_1_2B	Liquid AI LFM 2.5 1.2B general-purpose hybrid language model.	XNNPACK, MLX	8da4w, int4	v0.8.0
models.llm.LFM2_5_2_6B	Liquid AI LFM 2.5 2.6B hybrid language model post-trained for instruction following and agentic workloads.	XNNPACK, MLX	8da4w, int4	v0.11.0
models.llm.LFM2_5_350M	Liquid AI LFM 2.5 350M ultra-compact hybrid language model.	XNNPACK, MLX	8da4w, int4	v0.8.1
models.llm.LFM2_5_VL_450M	Liquid AI LFM 2.5 VL 450M lightweight multimodal vision-language model. Inputs: text, image.	XNNPACK, MLX, Vulkan	8da4w, int4	v0.8.4
models.llm.LFM2_5_VL_1_6B	Liquid AI LFM 2.5 VL 1.6B high-capacity vision-language model. Inputs: text, image.	XNNPACK, MLX, Vulkan	int8, 8da4w, int4	v0.8.4
models.llm.LFM3_1_7B	Liquid AI LFM 3 1.7B hybrid language model (gated short convolutions, sliding-window and global attention).	XNNPACK, MLX	int8, 8da4w, int4	next
models.llm.LFM3_VL_2B	Liquid AI LFM 3 VL 2B vision-language model on the LFM 3 backbone with a hybrid convolution/attention vision encoder. Inputs: text, image.	XNNPACK, MLX	int8, 8da4w, int4	next
models.llm.LFM3_AUDIO_2B	Liquid AI LFM 3 Audio 2B speech-understanding model on the LFM 3 backbone with a causal audio encoder (16 kHz, up to 30 s per encoder call). Inputs: text, audio.	XNNPACK, MLX	int8, 8da4w, int4	next
models.llm.BIELIK_V3_1_5B	Bielik v3 1.5B bilingual Polish & English language model, developed by SpeakLeash.	XNNPACK	8da4w	v0.9.0
models.llm.LLAMA3_2_1B	Meta Llama 3.2 1B lightweight instruction-tuned multilingual model.	XNNPACK, MLX	int4, spinquant	v0.1.0
models.llm.LLAMA3_2_3B	Meta Llama 3.2 3B instruction-tuned multilingual language model.	XNNPACK, MLX	int4, spinquant	v0.1.0
models.llm.SMOLLM2_135M	Hugging Face SmolLM2 135M ultra-compact language model.	XNNPACK, MLX	int8, 8da8w	v0.4.1
models.llm.SMOLLM2_360M	Hugging Face SmolLM2 360M compact instruction-tuned model.	XNNPACK, MLX	int8, 8da8w	v0.4.1
models.llm.SMOLLM2_1_7B	Hugging Face SmolLM2 1.7B language model trained on curated educational, synthetic, and web data.	XNNPACK, MLX	int8, 8da8w	v0.4.1
models.llm.HAMMER2_1_0_5B	Hammer 2.1 0.5B specialized function-calling model.	XNNPACK, MLX	8da4w, int4	v0.4.1
models.llm.HAMMER2_1_1_5B	Hammer 2.1 1.5B function-calling language model.	XNNPACK, MLX	8da4w, int4	v0.4.1
models.llm.HAMMER2_1_3B	Hammer 2.1 3B high-capacity function-calling model.	XNNPACK, MLX	8da4w, int4	v0.4.1
models.llm.PHI4_MINI	Microsoft Phi-4 Mini 3.8B high-density reasoning model.	XNNPACK, MLX	8da4w, int4	v0.4.1
models.llm.QWEN2_5_0_5B	Alibaba Qwen 2.5 0.5B ultra-lightweight multilingual model.	XNNPACK, MLX	8da4w, int4	v0.4.1
models.llm.QWEN2_5_1_5B	Alibaba Qwen 2.5 1.5B multilingual instruction model.	XNNPACK, MLX	8da4w, int4	v0.4.1
models.llm.QWEN2_5_3B	Alibaba Qwen 2.5 3B high-capability multilingual model.	XNNPACK, MLX	8da4w, int4	v0.4.1
models.llm.QWEN3_0_6B	Alibaba Qwen 3 0.6B next-generation compact language model.	XNNPACK, MLX	int8, 8da4w	v0.4.1
models.llm.QWEN3_1_7B	Alibaba Qwen 3 1.7B next-generation multilingual language model.	XNNPACK, MLX	8da4w, int4	v0.4.1
models.llm.QWEN3_4B	Alibaba Qwen 3 4B high-capacity generative model.	XNNPACK, MLX	8da4w, int4	v0.4.1
models.llm.GEMMA4_E2B	Google Gemma 4 E2B generative language model.	XNNPACK, MLX, Vulkan	8da4w, int4	v0.9.1
models.llm.GEMMA4_OMNI_E2B	Google Gemma 4 E2B multimodal (omni) language model supporting text, image, and audio inputs. Inputs: text, image, audio.	XNNPACK, MLX, Vulkan	8da4w, int4	v0.9.1

Text Embeddings

Guide · Benchmarks

Model	Description	Backends	Precisions	Added in
models.textEmbeddings.ALL_MINILM_L6_V2	Compact 384-dimensional sentence transformer mapping text to a dense vector space.	XNNPACK, Core ML, Vulkan	fp16, int8	v0.4.1
models.textEmbeddings.ALL_MPNET_BASE_V2	High-quality 768-dimensional sentence transformer model based on MPNet.	XNNPACK, Vulkan	fp16, int8, 8da4w	v0.4.1
models.textEmbeddings.MULTI_QA_MINILM_L6_COS_V1	384-dimensional sentence transformer fine-tuned specifically for semantic QA matching using cosine similarity.	XNNPACK, Core ML, Vulkan	fp16, int8	v0.4.1
models.textEmbeddings.MULTI_QA_MPNET_BASE_DOT_V1	768-dimensional sentence transformer fine-tuned specifically for question-answering matching using dot product distance.	XNNPACK, Vulkan	fp16, int8, 8da4w	v0.4.1
models.textEmbeddings.PARAPHRASE_MULTILINGUAL_MINILM_L12_V2	384-dimensional sentence transformer supporting 50+ languages for cross-lingual semantic similarity.	XNNPACK, Core ML, Vulkan	fp16, int8, 8da4w	v0.9.0
models.textEmbeddings.DISTILUSE_BASE_MULTILINGUAL_CASED_V2	Multilingual sentence transformer supporting 50+ languages, based on distilled Universal Sentence Encoder (512-dim output).	XNNPACK, Core ML, MLX, Vulkan	fp16, int8, 8da4w	v0.9.0
models.textEmbeddings.CLIP_VIT_BASE_PATCH32_TEXT	CLIP text encoder (ViT-B/32) mapping text queries into a 512-dimensional joint text-image embedding space.	XNNPACK, Core ML, Vulkan	fp16, int8	v0.3.4
models.textEmbeddings.LFM2_5_EMBEDDING_350M	Liquid AI LFM 2.5 350M parameter embedding model for asymmetric search and retrieval tasks.	XNNPACK, MLX	8da4w, int4	v0.10.0

ColBERT Retrieval

Guide · Benchmarks

Model	Description	Backends	Precisions	Added in
models.colbertEmbeddings.LFM2_5_COLBERT_350M	Liquid AI LFM 2.5 350M ColBERT model: 128-dimensional token vectors, multilingual (11 languages), 32-token queries and documents of up to 512 tokens.	XNNPACK, MLX	8da4w, int4	v0.11.0

Privacy Filter (PII detection)

Guide · Benchmarks

Model	Description	Backends	Precisions	Added in
models.privacyFilter.OPENAI	OpenAI-style detector covering 8 common PII types (name, email, phone, address, and similar).	XNNPACK, MLX	8da4w, int4	v0.9.0
models.privacyFilter.NEMOTRON	Nemotron-based detector covering 55 fine-grained PII types.	XNNPACK, MLX	int8, 8da4w	v0.9.0

Translation

Guide

Model	Description	Backends	Precisions	Added in
models.translation.OPUS_MT_EN_DE	Opus-MT English to German.	XNNPACK	int8	next
models.translation.OPUS_MT_DE_EN	Opus-MT German to English.	XNNPACK	int8	next
models.translation.OPUS_MT_EN_ES	Opus-MT English to Spanish.	XNNPACK	int8	next
models.translation.OPUS_MT_ES_EN	Opus-MT Spanish to English.	XNNPACK	int8	next
models.translation.OPUS_MT_EN_FR	Opus-MT English to French.	XNNPACK	int8	next
models.translation.OPUS_MT_FR_EN	Opus-MT French to English.	XNNPACK	int8	next
models.translation.EUROLLM_1_7B	EuroLLM 1.7B Instruct: any two of its 35 languages with one model.	XNNPACK, MLX	int8, 8da8w	next
models.translation.HY_MT2_1_8B	Hy-MT2 1.8B: any two of its 38 languages and dialects with one model.	XNNPACK, MLX	int8, 8da4w	next

Speech to Text

Guide · Benchmarks

Model	Description	Backends	Precisions	Added in
models.speechToText.WHISPER	OpenAI Whisper automatic speech recognition model family with integrated Voice Activity Detection. Variants: TINY, BASE, SMALL, EN.TINY, EN.BASE, EN.SMALL.	XNNPACK, Core ML, MLX, Vulkan	fp16, bf16, int8	v0.3.0
models.speechToText.QWEN3_ASR	Qwen3-ASR speech recognition with integrated Voice Activity Detection.	XNNPACK, MLX	8da4w, coreml_int4	next

Text to Speech

Guide · Benchmarks

Model	Description	Backends	Precisions	Added in
models.textToSpeech.SUPERTONIC	Supertonic 3 multilingual flow-matching Text-to-Speech model.	XNNPACK, MLX, Vulkan	fp32, fp16, int8	v0.10.0
models.textToSpeech.KOKORO	Kokoro — a lightweight phoneme-driven Text-to-Speech model. Variants: EN_US, EN_GB, ES, FR, IT, PT, HI, PL, DE, ZH.	XNNPACK, Core ML	fp32	v0.7.0

Voice Activity Detection

Guide · Benchmarks

Model	Description	Backends	Precisions	Added in
models.voiceActivityDetection.FSMN_VAD	Feedforward Sequential Memory Network (FSMN) Voice Activity Detection model.	XNNPACK	fp32	v0.6.0

Image Classification

Guide · Benchmarks

Model	Description	Backends	Precisions	Added in
models.classification.EFFICIENTNET_V2_S	EfficientNetV2-S image classification model pre-trained on ImageNet-1k (1000 categories, see IMAGENET1K_LABELS).	XNNPACK, Core ML	fp16, int8	v0.2.0

Object Detection

Guide · Benchmarks

Model	Description	Backends	Precisions	Added in
models.objectDetection.SSDLITE320_MOBILENET_V3_LARGE	SSDLite object detector with MobileNetV3-Large backbone trained on COCO (see COCO_CLASSES) at 320x320 resolution.	XNNPACK, Core ML	fp32, fp16	v0.2.0
models.objectDetection.RFDETR_NANO	RF-DETR (Roboflow Detection Transformer) Nano variant trained on COCO (see COCO_CLASSES).	XNNPACK, Core ML, Vulkan	fp16, int8	v0.8.0
models.objectDetection.YOLO26	Ultralytics YOLO26 real-time object detection models trained on COCO (80 classes, see COCO_CLASSES_YOLO). Variants: NANO, SMALL, MEDIUM, LARGE, XLARGE (each SIZE_384, SIZE_512, SIZE_640).	XNNPACK, Core ML, Vulkan	fp16, int8	v0.8.0

Instance Segmentation

Guide · Benchmarks

Model	Description	Backends	Precisions	Added in
models.instanceSegmentation.FASTSAM	Fast Segment Anything Model (FastSAM) for promptable or global object instance mask segmentation. Variants: S, X.	XNNPACK, Core ML, Vulkan	fp16, int8	v0.9.0
models.instanceSegmentation.RFDETR_NANO	RF-DETR (Roboflow Detection Transformer) Nano instance segmentation model predicting COCO class masks and bounding boxes (see COCO_CLASSES).	XNNPACK, Core ML, Vulkan	fp16, int8	v0.8.0
models.instanceSegmentation.YOLO26	YOLO26 instance segmentation models predicting COCO class instance masks and bounding boxes (see COCO_CLASSES_YOLO). Variants: NANO, SMALL, MEDIUM, LARGE, XLARGE (each SIZE_384, SIZE_512, SIZE_640).	XNNPACK, Core ML, Vulkan	fp16, int8	v0.8.0

Semantic Segmentation

Guide · Benchmarks

Model	Description	Backends	Precisions	Added in
models.semanticSegmentation.SELFIE_SEGMENTATION	Lightweight portrait selfie segmentation model for real-time person vs background separation.	XNNPACK, Core ML	fp32, fp16	v0.8.0
models.semanticSegmentation.SELFIE_SEGMENTATION_LANDSCAPE	MediaPipe Selfie Segmentation, landscape orientation.	XNNPACK, Core ML	fp32, fp16	v0.10.0
models.semanticSegmentation.U2NETP	Lightweight U²-Netp foreground segmentation at 320x320 for objects and people.	XNNPACK, Core ML, Vulkan	fp32, fp16, int8	next
models.semanticSegmentation.LRASPP_MOBILENET_V3_LARGE	Lite R-ASPP semantic segmentation model with MobileNetV3-Large backbone (21 classes, see PASCAL_VOC_LABELS).	XNNPACK, Core ML	fp16, int8	v0.8.0
models.semanticSegmentation.DEEPLAB_V3_RESNET50	DeepLabV3 semantic segmentation model with ResNet-50 backbone (21 classes, see PASCAL_VOC_LABELS).	XNNPACK, Core ML	fp16, int8	v0.4.1
models.semanticSegmentation.DEEPLAB_V3_RESNET101	DeepLabV3 semantic segmentation model with ResNet-101 backbone (21 classes, see PASCAL_VOC_LABELS).	XNNPACK, Core ML	fp16, int8	v0.8.0
models.semanticSegmentation.DEEPLAB_V3_MOBILENET_V3_LARGE	DeepLabV3 semantic segmentation model with MobileNetV3-Large backbone (21 classes, see PASCAL_VOC_LABELS).	XNNPACK, Core ML	fp16, int8	v0.8.0
models.semanticSegmentation.FCN_RESNET50	Fully Convolutional Network (FCN) semantic segmentation model with ResNet-50 backbone (21 classes, see PASCAL_VOC_LABELS).	XNNPACK, Core ML	fp16, int8	v0.8.0
models.semanticSegmentation.FCN_RESNET101	Fully Convolutional Network (FCN) semantic segmentation model with ResNet-101 backbone (21 classes, see PASCAL_VOC_LABELS).	XNNPACK, Core ML	fp16, int8	v0.8.0

Pose & Keypoints

Guide · Benchmarks

Model	Description	Backends	Precisions	Added in
models.keypointDetection.BLAZEFACE	MediaPipe BlazeFace lightweight face detection and 6-point facial landmark locator (eyes, nose, mouth, ears, see BLAZEFACE_LANDMARKS).	XNNPACK, Vulkan	fp32, fp16	v0.10.0
models.keypointDetection.FACEMESH	MediaPipe Face Mesh: a dense 468-point 3-D mesh (see FACEMESH_LANDMARKS) and a face-presence score, regressed from a 192x192 crop of one face.	XNNPACK, Core ML	fp32, fp16	v0.11.0
models.keypointDetection.YOLO26_POSE	YOLO26 human pose estimation model predicting 17 COCO body keypoints (see COCO_LANDMARKS). Variants: SIZE_384, SIZE_512, SIZE_640.	XNNPACK, Core ML, Vulkan	fp16, int8	v0.9.0
models.keypointDetection.RFDETR_KEYPOINT	RF-DETR (Roboflow Detection Transformer) pose keypoint detector predicting 17 COCO body keypoints (see COCO_LANDMARKS).	XNNPACK, Core ML, Vulkan	fp16, int8	v0.9.2

Depth Estimation

Guide · Benchmarks

Model	Description	Backends	Precisions	Added in
models.depthEstimation.DEPTH_ANYTHING_V2_SMALL	Depth Anything V2 Small with a ViT-S/14 backbone.	XNNPACK, Core ML, Vulkan	fp32, fp16, int8	v0.11.0
models.depthEstimation.DEPTH_ANYTHING_V3_SMALL	Depth Anything 3 Small, exported for single-image relative depth.	XNNPACK, Core ML, Vulkan	fp32, fp16, int8	v0.11.0
models.depthEstimation.ZIPDEPTH_BASE	ZipDepth Base NPU with a 384 × 384 input for low-latency depth estimation.	XNNPACK, Core ML, Vulkan	fp32, fp16	v0.11.0

OCR

Guide · Benchmarks

Model	Description	Backends	Precisions	Added in
models.ocr.PADDLE	PP-OCRv6 small — DBNet detector plus an SVTR recognizer, one model for every language.	XNNPACK, Core ML, Vulkan	fp16, int8	v0.10.0

Image Embeddings

Guide · Benchmarks

Model	Description	Backends	Precisions	Added in
models.imageEmbeddings.CLIP_VIT_BASE_PATCH32	CLIP vision encoder (ViT-B/32) mapping images into a 512-dimensional shared text-image space.	XNNPACK, Core ML, MLX, Vulkan	fp16, int8	v0.5.0

Image Inpainting

Guide · Benchmarks

Model	Description	Backends	Precisions	Added in
models.inpainting.MIGAN	MI-GAN, a mobile inpainting GAN trained on Places2 (scenes) and FFHQ (faces). Variants: SIZE_512, SIZE_256, FACES_256.	XNNPACK, Core ML, Vulkan	fp16	v0.11.0

Text to Image

Guide · Benchmarks

Model	Description	Backends	Precisions	Added in
models.textToImage.SDXS_512_DREAMSHAPER	Ultra-fast SDXS (Stable Diffusion eXtreme Speed) 512x512 text-to-image generation model based on DreamShaper.	XNNPACK, Core ML, Vulkan	fp16, int8	v0.10.0

Style Transfer

Guide · Benchmarks

Model	Description	Backends	Precisions	Added in
models.styleTransfer.CANDY	Fast neural style transfer model generating a vibrant, artistic "Candy" style effect.	XNNPACK, Core ML, Vulkan	fp16, int8	v0.2.0
models.styleTransfer.MOSAIC	Fast neural style transfer model applying a classic tile mosaic artistic pattern.	XNNPACK, Core ML, Vulkan	fp16, int8	v0.2.0
models.styleTransfer.RAIN_PRINCESS	Fast neural style transfer model applying a painterly "Rain Princess" oil painting aesthetic.	XNNPACK, Core ML, Vulkan	fp16, int8	v0.2.0
models.styleTransfer.UDNIE	Fast neural style transfer model applying Francis Picabia's "Udnie" abstract art style.	XNNPACK, Core ML, Vulkan	fp16, int8	v0.2.0