Locai Labs researches how to make advanced AI models practical on local and edge hardware.
Our inference research includes structured pruning, post-training quantisation and GPU-specific kernel optimisation. These techniques reduce the memory and compute needed to serve a model at useful speed.
Our training research focuses on continual learning: adapting pretrained models to new private data while limiting catastrophic forgetting. Catastrophic forgetting happens when a model learns something new but loses capabilities it previously had.
Two named parts of this research are SPACE, our specialisation pruning approach for compressing mixture-of-experts models, and Forget-Me-Not, our continual-learning work for retaining existing capability during further training.