- Category
- AI Tools
- Rank
- No. 929Tools index
Previous survey · No. 919 ·
- Pricing
- Open Source
- Type
- TOOL
- Builder
- google-ai-edge
- GitHub
- 41 stars
- Latest release
- v0.1.1
- Date
About
CLI for the LiteRT (TFLite) on-device runtime — convert, quantize, compile, run, benchmark and visualize models across CPU, GPU, and NPU.
What it does
LiteRT CLI gives edge model developers one command surface for managing models from acquisition through device testing. It keeps downloaded models in a local cache, resolves reusable model references, installs command-specific dependencies when needed, and can hand language-model work to the separate LiteRT-LM executable.
Why it's ranked here
The broad workflow coverage is useful, especially when comparing model variants on desktop and connected Android hardware. Command modules are imported only when their command is invoked, and optional feature packages install on first use. The preview warning matters, however: several platform combinations remain unavailable, and some advanced paths depend on external Google tooling or access programs.
What's good
The base installation stays smaller because conversion, compilation, quantization, visualization, language-model, and audio dependencies are separated by feature. Model references make cached downloads reusable across commands. The tool also accepts arrays, binary tensors, NumPy files, and common image formats, which reduces setup for basic inference experiments.
Tradeoffs
This is explicitly an early preview with possible bugs and limited support. Compilation requires Linux and a recent Clang toolchain, while Windows lacks compilation and conversion. NPU compilation currently targets Qualcomm hardware; MediaTek and Google Tensor support are described as forthcoming. Image preprocessing uses generic assumptions that may not match a model's required normalization.
How to use it well
Use it for repeatable edge deployment experiments: fetch a model, create optimized variants, then run comparable measurements on desktop or Android targets. It best suits developers already working with LiteRT and related Google AI Edge components. Keep model training, dataset preparation, and production application integration in separate tools and workflows.
Technical notes+
pyproject.toml defines a setuptools Python package, the litert console entry point, core dependencies, and feature extras. litert_cli/litert.py maps twelve commands to modules and imports each module only when that command is requested. litert_cli/core/deps.py checks installed distributions and invokes uv or pip for missing extras. litert_cli/commands/lm.py forwards arguments to the external litert-lm process. litert_cli/core/inputs.py handles tensor literals, NumPy arrays, raw binaries, and images, while litert_cli/litert_test.py and litert_cli/litert_help_test.py test command discovery and help loading.
Observed
- License
- Apache-2.0
- Primary language
- Python
- Interface
- Command-line interface exposed through the litert console command
- Packaging
- Setuptools package installable from PyPI with pip or uv, or as an editable local clone
- Python requirement
- Python 3.10 or newer; repository instructions say verified with Python 3.13
- Verified platforms
- Linux, Apple Silicon macOS, Windows, and Android, with command and accelerator limitations
- Test structure
- Command discovery and help-loading tests are colocated under the litert_cli package
Read from README.md, pyproject.toml, litert_cli/litert.py, tools/build_wheels.py, litert_cli/__init__.py, litert_cli/litert_test.py, litert_cli/litert_help_test.py, litert_cli/core/deps.py, litert_cli/core/utils.py, litert_cli/commands/lm.py, litert_cli/core/inputs.py, litert_cli/core/models.py, litert_cli/models/base.py, litert_cli/commands/list.py, litert_cli/commands/clean.py.
What it can do
Convert models to LiteRT format
Machine learning models → LiteRT (TFLite) format models
Quantize neural network models
LiteRT models → Quantized models with reduced precision
Compile models for specific hardware
LiteRT models → Hardware-optimized compiled models
Run inference on models
LiteRT models and input data → Model predictions/inference results
Benchmark model performance
LiteRT models → Performance metrics and timing data
Visualize model architecture
LiteRT models → Model structure visualizations
Execute models on CPU
LiteRT models → CPU-based inference results
Execute models on GPU/NPU
LiteRT models → Hardware-accelerated inference results
Tags
Tech Stack
Media
Comments (0)
No comments yet
Editorially curated, with community endorsements as a secondary signal. Corrections welcome.