compress your models.
keep your accuracy.
FasterAI prunes, quantizes and distills PyTorch models to the target you choose (speed, size, or a specific chip), under an accuracy floor you set.
Apache 2.0 · ★ 265 on GitHub · maintained since 2021
real gains, on your hardware.
faster inference
smaller model
less energy
Example measured run: 3.4× faster · −88% size · −71% energy · −0.9 pt accuracy. see the full report →
you pick the outcome.
we pick the technique.
A business goal
optimize(model, sample,
target='speed')Or 'size', or 'energy'. Say what you need; the pipeline is chosen for you.
A specific chip
optimize(model, sample,
target='nvidia-t4')
# or 'stm32-h7', 'jetson-orin'…From datacenter GPUs to microcontrollers: the same call, tuned to the hardware.
Accuracy is a hard constraint, not a hope
optimize(model, sample,
target='speed',
constraints=Constraints(
min_accuracy=90))Set a floor. Compression never trades below it.
two ways to optimize.
Do it yourself
Everything you need to compress models on your own, free and open.
Apache 2.0 · ★ 265 · since 2021
Read the quickstartWork with us
Fixed-scope audit first: you know the gains before committing to anything.
Book a callmake your models faster this week.
Engineers start with the library. Teams start with an audit.