Every model in litert-community, with its source model, its on-device benchmarks and its conversion recipe in litert-samples, where these exist.
Reading the model list from Hugging Face…
Could not load the model list from Hugging Face (). Reload to try again, or see the models at huggingface.co/litert-community.
This page stopped on an error (). The models are at huggingface.co/litert-community.
Benchmarks could not be loaded, so models are listed without them. They are also on the leaderboard in litert-samples.
Recipe links could not be loaded from litert-samples. The recipes are at litert-samples/models.
No models match these filters.
LiteRT CLI downloads a model from the org and benchmarks it on a phone over adb (--android), on the Mac it runs on (--desktop), or on a lab phone in Google Cloud's Developer Device Platform (--ddp, a session billed to your project); --cpu, --gpu and --npu pick the accelerator:
pip install litert-cli-nightly
litert download litert-community/MobileNet-v2 --file "*.tflite" --output mobilenet
litert benchmark mobilenet/mobilenet_v2.tflite --desktop --cpu
litert benchmark mobilenet/mobilenet_v2.tflite --android --gpu
litert benchmark mobilenet/mobilenet_v2.tflite --ddp --device pa3q-35 --gcp-project your-project-idThe codelab under benchmark/developer_device_platform in litert-samples benchmarks a model on a lab phone from a Colab runtime, with nothing installed on your machine. Its two notebooks: litert_cli_benchmark.ipynb runs litert benchmark --ddp on a .tflite file and on a .litertlm bundle, and ddp_cli_benchmark.ipynb calls the Developer Device Platform CLI directly, for several devices at once or another benchmark binary.
Each model's details carry the download and desktop lines with its own file name, plus litert run for a .tflite file and litert lm benchmark and litert lm run for a .litertlm bundle. The CLI's README lists every form, a .litertlm bundle included. The benchmarks on this page are rows of the leaderboard in litert-samples; litert-samples/benchmark holds the drivers that made them, the iPhone app, and the README on how a row is made.
Models: the Hugging Face API, read when the page loads; repos that hold only a README are left out. Source model: the base_model field of the repo's card. Benchmarks: the data file of the leaderboard in litert-samples, read when the page loads. Recipes: the model list in litert-samples/models/README.md, read when the page loads. A benchmark is one model file on one platform, device and accelerator, at the newest runtime version measured, and a comparison holds only within one platform, device, accelerator and task. Task is the repo's pipeline tag, and Format the .tflite, .litertlm and .task files a repo holds. Family is the leading letters of the source model's name (of the repo's name when the card names none), so gemma-3-1b-it and Gemma3-1B-IT file together; Size is the parameter count in the repo's name or its source model's name, such as 0.6B, 4B or 270M, and a model whose names carry none is under "Size not in the name". The CLI lines in a model's details take the file with the shortest name of each format.