Navigation
Offerings

Our platform is deployed across government, defense, and enterprise organizations that require complete data sovereignty and AI capability without cloud dependencies.

↳ Learn more about CASTLE
Quick Links
EDGE  /  CANADA'S FOUNDATION MODEL

Canada's
Foundation Model

Built from scratch. Zero foreign model weights. Every parameter trained on Canadian hardware, under Canadian oversight, answerable to Canadian law.

Inspect the Model
EDGE-1.3B / TRAINING LIVE
1.34B
Parameters
5.95B
Training Tokens
0
Foreign Weights
100%
Canadian
step  6,200/100,000
loss  6.41 ↓   lr 1.49e-04
480 tok/s  •  ETA 57h
Hardware: Apple Silicon, Canada
Cloud used: None
Training Progress 6%

Every Major AI Nation Has
a Foundation Model. Canada Now Does Too.

Every major AI-capable nation has a model it calls its own. The United States has GPT and Claude. France has Mistral. China has Qwen and DeepSeek. Canada — home of neural networks, the Vector Institute, and some of the world's most cited AI researchers — has had nothing it can point to and call its own foundation. Until now.

1.34B
Parameters training now
0
Foreign model weights
100%
Canadian hardware & oversight

Three Generations.
One Canadian Foundation.

Edge is a model family, not a one-off experiment. Each generation improves on the last — more data, better architecture, greater capability. All trained from scratch. All Canadian.

Live
Edge-Medium
85M Parameters
Architecture proof. Trained on WikiText-103. Runs on any laptop. Fully auditable — inspect every weight and training decision.
View Weights ↗
Training
Edge-1.3B
1.34B Parameters
Flagship. Training on 5.95B tokens of FineWeb on Canadian hardware. Powers CASTLE inference, Mountie, and WARDEN when complete.
Follow Progress ↗
Planned 2026
Edge-7B
7B Parameters
Production scale. Modern architecture (RoPE + SwiGLU + GQA). Designed to compete with international foundation models on quality benchmarks.

Built Transparent.
Designed Sovereign.

Every architectural decision, every hyperparameter, every training step is documented and auditable. This is not a black box. It is the opposite of a black box.

24 Layers
Transformer depth calibrated for 1.3B parameter scale.
2048 Dimensions
Model width providing capacity for complex representations.
16 Attention Heads
Parallel attention enabling rich contextual understanding.
BPE 32K Vocab
Custom tokenizer trained on FineWeb data. Efficient, sovereign.
1024 Token Context
Processing window for document-length inputs.
MLX Native
Apple Silicon Metal GPU. No CUDA. No cloud compute. Full sovereign training.
Training Specification
DatasetFineWeb-Edu 5.95B tokens
HardwareApple Silicon, Canada
FrameworkMLX (Metal GPU)
Learning rate1.5e-4 cosine
Warmup2,000 steps
Target steps100,000+
Batch size1 × 1024 tokens
Cloud usedNone
Foreign weightsNone
ECOSYSTEM

Edge Is the Foundation.
CASTLE Is the Infrastructure.

Edge doesn't stand alone. It is the model layer of a complete Canadian AI stack — from the weights to the API to the consumer product to the defence capability.

The Model
Edge
Canadian-built foundation. Zero foreign weights. Fully auditable. The bedrock everything else runs on.
The Platform
CASTLE
OpenAI-compatible sovereign inference API. Echo routes queries to Edge. Zero external dependencies.
The Product
Mountie
Canadian government services AI. Powered by Edge on CASTLE. Sovereign from model to answer.
The Defence
WARDEN
Off-grid AI over UHF radio. Edge inference without internet. Zero cloud. Works when everything else fails.

EDGE POWERS ALL OF THEM  •  ZERO EXTERNAL DEPENDENCIES

Two Ways to Access Edge

Inspect the Model

For researchers and developers. The weights, architecture, training code, and full methodology are public. Inspect every parameter, every training decision, every checkpoint.

Open on HuggingFace ↗
Deploy via CASTLE

For organizations requiring sovereign deployment. Edge runs inside CASTLE — OpenAI-compatible API, your hardware, your data, your rules. Zero external dependencies.

AXe
CONTACT / DEMO REQUEST + PARTNERSHIP INQUIRY

Interested in solving your problems with AXe software?