Research
Open research on sovereign AI.
We publish our findings on efficient architectures, privacy-preserving inference, and bilingual evaluation. The AI community benefits when knowledge is shared.
Sparse Attention Mechanisms for 50% Inference Speedup
We introduce a learned sparsity pattern that retains 99.2% of dense-attention quality while halving wall-clock latency on long contexts.
Privacy-Preserving Inference Under PIPEDA
A practical architecture for zero-retention LLM serving in regulated Canadian environments, including a formal threat model.
QC-Bench: A Benchmark for Quebec French Language Models
We release QC-Bench, the first comprehensive benchmark for Québécois French covering 12 tasks across joual, formal register, and code-switching.
Loss-Aware INT4 Quantization for Edge Deployment
A new quantization scheme that bounds per-layer activation error, enabling 3B parameter models to run at full quality on 8GB consumer GPUs.
Constitutional Alignment in a Bilingual Setting
Reward-model transfer between English and Quebec French preference data, with ablations on cross-lingual harm taxonomies.
Carbon Accounting for Hydroelectric LLM Training
A transparent methodology for reporting embodied and operational carbon when training on Quebec's grid, with full release of our 70B run.