Inference
Efficient inference.
Structured pruning, post-training quantisation, and GPU-specific kernels so large open-weight models serve at interactive latency in GPU constrained environments.
Pruning algorithms
Removing low-saliency parameters under a task-retention constraint.
Dynamic quantisation strategies
Adaptive reduction of weight and activation precision to match workload and hardware.
Kernel optimisation
AutoKernel and LLM-generated kernels tuned for the target GPU.
Training
Continual learning.
Continual adaptation of pretrained weights to private data, with methods that limit catastrophic forgetting. No full pretraining run required.
Sample efficient continual learning
Adapting from limited new data without a full pretraining run.
Catastrophic forgetting
Methods including metaplasticity that bound loss of prior capability during sequential updates.
Efficient parametric learning
Parameter-efficient updates instead of training a model from scratch.
SPACE™: Specialisation Pruning
Algorithm for Compression of Experts
Structured pruning is normally objective-agnostic: minimise divergence from the base model across its entire capability surface. SPACE conditions the objective on a target distribution instead. Score the specialised subnetworks for saliency to the capabilities you need, prune outside them, and trade breadth you were never going to use for a parameter count that serves at interactive latency in GPU constrained environments.
Original model
Large generalist model
Specialisations identified
Scoring reveals the specialised subnetworks within the model
Model compressed
Irrelevant parameters are pruned, leaving a compressed, specialised model
Compress a state-of-the-art model far enough and the data centre becomes the box in the corner of your office.
Our model series
Juno
July 2026
Capability-targeted compression of open-weight bases with SPACE: structured pruning and post-training quantisation under a task-retention constraint for GPU constrained environments.
Juno-NJupiter-N
April 2026
Post-train of NVIDIA Nemotron 3 Super 120B for British context, Welsh language, UK cultural grounding and stronger agentic performance. Forget-Me-Not keeps the base capabilities intact.
Jupiter-G
April 2026
The same Locai post-training recipe on Google Gemma-4, lifting instruction following, safety and code without regressing general knowledge.
L1
November 2025
Continual post-training with Forget-Me-Not: experience replay and self-generated preference data. Improves preference alignment and low-resource language performance without regressing STEM capability on the base.
Intellectual Property
This patent covers systems and computer-implemented methods that consolidate experiences of a machine learning model in order to improve model performance for continual learning.
This patent introduces novel methods for fine-tuning machine learning models while preserving previously learned knowledge. The techniques address the critical challenge of catastrophic forgetting that occurs during incremental learning.
This foundational patent covers autonomous decision-making capabilities in AI agents while maintaining safety and control through sophisticated guardrails.
View patent →This patent describes breakthrough techniques for enhancing speech recognition performance, particularly in challenging acoustic environments.
View patent →This patent covers innovative approaches to automatically classify and route telecommunication calls in real-time using advanced machine learning techniques.
View patent →