Skip to content
Projects

john-rocky/LiteRT-Models has 29 GitHub stars and 4 forks. GitHot measured 0 stars in 24 hours. Last checked 2026-11-16 03:10 UTC. Last push 2026-11-13.

How these counts are measured

Measured star growth

in 24h
0
in 7 days
Collecting history
in 15 days
Collecting history
in 1 month
Collecting history

Net star changes over overlapping periods, not daily totals. Missing measurements are left blank. Stars measure attention, not software quality.

About

Converted TFLite Model Zoo for Android with LiteRT CompiledModel GPU acceleration

Machine Learning

README.md

Original

LiteRT-Models

This repo has more than 120 models converted to run on Android with LiteRT, Google's on-device ML runtime. LiteRT is the new name of TensorFlow Lite. Most models come with input and output shapes, preprocessing, the conversion script and a standalone Kotlin sample app.

Most of the .tflite models run every op on the phone's GPU through CompiledModel. For example, DINOv2 ViT-S/14 at 448² runs in 53.6 ms on a Galaxy S26 GPU. The chat models run on LiteRT-LM, on a phone or a Mac.

If you like this repository, please give it a star.

Run a model

Download a .tflite from MODELS.md, put it in app/src/main/assets/, and add the dependency from Google Maven:

implementation("com.google.ai.edge.litert:litert:2.1.3")
val options = CompiledModel.Options(Accelerator.GPU)
val model = CompiledModel.create(context.assets, "model.tflite", options, null)
val inputs = model.createInputBuffers()
inputs[0].writeFloat(floatArray)
val outputs = model.run(inputs)
val result = outputs[0].readFloat()

Each model's section in MODELS.md says how to prepare floatArray. Reading the output waits for the GPU, so time the run and the read together. Models over about 150 MB load from a file path instead of the APK. From LiteRT 2.3.0, the GPU also needs com.google.ai.edge.litert:litert-gpu at the same version as litert.

What's here

Task Models
Detection and tracking YOLO11n, YOLO26n, RF-DETR Nano, RT-DETRv2-S, D-FINE-S, SSDLite320 MobileNetV3, YOLO + DeepSORT, Ultra-Fast-Lane-Detection, TwinLiteNet, M-LSD-tiny, YuNet, RTMPose-Face, SINet-V2, DM-Count
Segmentation and cutouts PIDNet-S, YOLACT-ResNet50, MobileSAM, SAM 2.1, EdgeTAM, EdgeTAM Video, CLIPSeg, Cloth Segmentation (U²-Net), RMBG-1.4, ormbg, DIS, MODNet, WeVisOne-VL-2B
Pose and face YOLO26n-pose, RTMPose-s, RTMW-m, RTMPose-Hand, RTMPose-Animal, 6DRepNet, L2CS-Net, 3DDFA_V2, BiSeNet, HSEmotion, Silent-Face
Image enhancement and editing EDSR, Real-ESRGAN x4v3, GFPGAN v1.4, DehazeFormer-MCT, CPGA-Net, NAFNet, LaMa-Dilated, MI-GAN, Fast Neural Style, U²-Net Portrait, DewarpNet
Understanding images and video Vision-RWKV, PlantNet-300K, CLIP ViT-B/32, Places365 ResNet18, RAM++, NIMA, XFeat, UniSal, DINOv2 ViT-S/14, PP-OCRv5, MoViNet-A0
Depth and 3D Idol, MoGe-2 ViT-S, Depth Anything 3 ViT-S, Metric3D v2 ViT-S, TIPSv2-B/14 DPT, DSINE
Image generation iris-3b Turbo, Bonsai Image 4B
Speech recognition Parakeet, Whisper-tiny, wav2vec2-CTC, Qwen3-ASR-1.7B, Qwen3-ASR-1.7B-15s, Fun-ASR-Nano-2512, Confucius4-R2T2
Text-to-speech Kokoro-82M, Matcha-TTS, EMA Lightning, Dia2-1B, VibeVoice-Realtime-0.5B, Pocket TTS, KittenTTS nano 0.8, Inflect-Nano-v2, Sopro v2 turbo, Audio8-TTS-Preview-0.6b, Ming-3.0-tts-0.5B
Voice assistant, speaker labels and noise removal pyannote 3.1 stack, Nemotron-3-Diarization, CMGAN, RE-USE, Whisper + SmolLM2 + Kokoro pipeline
Audio and music wav2vec2 Keyword Spotting, PANNs CNN14 Audio Tagging, DAC 16kHz, Mimi, TIGER-DnR, CREPE, Basic Pitch
Chat and text generation RWKV-7 World 0.1B, Falcon3-3B-Instruct, Llama-3.2-3B-Instruct, Ministral-3-3B-Instruct-2512, SmolLM3-3B, Qwen3.8-27B
Vision-language SmolVLM-256M, SmolVLA
Embeddings, reranking, entity extraction and text classification Qwen3-Embedding-0.6B, Qwen3-Reranker-0.6B, GLiNER2.5 Small, GLiNER2.5 Multi, GLiFormer Large v1 — NER path, Laya Multilingual, ModernBERT-Ja-310M Decision, GLiNER2.5-Decide, GLiClass-Edge v3.0
Decision models: they answer yes-no, choice and score questions about text, JSON or a photo Julia-1, Phocinae-Largha-150M, Open Decision, Kev-0.8B, d1-omni-600M, d1-2B, d1-Omni-2B, d1-3B, clef-flash
Time-series forecasting Timestral-209M

MODELS.md lists every model with its speed and download.

Convert your own

To make a PyTorch model run fully on the GPU, start with the conversion guide. litert_gpu_toolkit/ has its rewrites as a Python package.

Coming from TensorFlow Lite?

The .tflite files still work. The package names changed: docs/litert-names.md maps old to new.

Projects using these models

  • PocketTTS-LiteRT runs Pocket TTS (Kyutai) on Android and builds on the pockettts/ module.
  • dpm, a tool for dancers to map the patterns they know, adapts the EdgeTAM video tracker in edgetam-video/ to anonymize video.
  • ImageSplitter, an Android app that splits a photo into numbered parts for large prints, upscales with real_esrgan_x4v3.tflite from this repo's v1 release.

License

MIT for the sample apps. kev/, julia1/, d1-3b/, clef-flash/ and wevisone/ carry their own Apache-2.0 LICENSE, and d1omni/ the LFM Open License v1.0, as it ports the model provider's code. Model licenses follow their original projects.