Efficient intelligence. Built for the real world.
Independent organization dedicated to developing and researching efficient and accessible AI.
Driven by a singular focus on
making powerful AI compatible with real-world hardware constraints,
we produce research, tools, and models exploring the practical limits of modern machine learning.
From small language models to efficient inference runtimes to aggressive quantization techniques, our projects investigate how the capabilities of modern AI can best be leveraged within the confines of memory, throughput, and computational power.
Efficient models
Inference & optimization
Model compression and quantisation
Accessible AI
Small does not mean simple
We believe in the fundamental importance of light-weight systems and architectures, and their ability to drastically broaden access to modern machine learning.
Lower requirements for memory, compute, bandwidth, and power enable a wider variety of deployment options, from personal experimentation to large-scale distributed systems.
We strive to publish our work openly whenever possible.
You'll find many of our models, experiments, datasets, and research artifacts in our
Hugging Face organization,
including various works-in-progress, ideas, benchmarks, and other detritus of the development process.
Efficient intelligence. Built for the real world.