Platform
CASTLE Crown Atlas Shield Tower Lens Echo Lab ↗ Edge Ontology Knox
Offerings

Production AI infrastructure for the rooms that can't use frontier APIs. Deployed on customer hardware. Audited end-to-end. Compounding on customer data.

↳ Explore the platform
Solutions
Government Defense Mountie Impact Studies
Company
About AXE Technologies CEO Vision Letter Trust & Custody Privacy News Innovation Lab Customer Portal Contact
Get Started

Talk to us about a CASTLE deployment in your environment.

↳ Request a demo
EDGECANADIAN IRON
FLAGSHIP1.19B PARAMS
CHINCHILLAV3 TRAINED
PURPOSE-BUILTNOT FINE-TUNED

EDGE

Canadian intelligence trained from scratch — custom tokenizer, custom GGUF, proprietary data. Not skins on open-source weights. Not fine-tuned costumes. Built from the ground up.

SCROLL

Models

Canadian Intelligence. Trained on Your Data.

Purpose-Built. Not Fine-Tuned.

Every AXE model is trained from the ground up — custom tokenizer, custom GGUF metadata, proprietary training data. These aren't skins on open-source weights. They're Canadian intelligence.

Casanova 2.0
70B Parameters
Role: Flagship — complex reasoning, research, strategic analysis
Speed: 8 tok/s on consumer hardware (64GB M-series)
Context: 32K tokens
Quantization: Q4_K_M GGUF
Training: AXE proprietary + curated open data
Geralt 2.0
11B Parameters
Role: Balanced — code generation, document analysis, reasoning
Speed: 25 tok/s
Context: 16K tokens
Quantization: Q4_K_M GGUF
Training: Code-heavy corpus + AXE interactions
Anakin 8B
8B Parameters
Role: Fast — chat, tool calling, real-time inference
Speed: 44 tok/s
Context: 8K tokens
Default: Echo routing
Training: Optimized for tool use

Micro Models. Millisecond Decisions.

Maestro 0.3B
Task classifier. Routes requests to the right model in <50ms.
FunctionAXE 0.3B
Tool router. Dispatches function calls with precision.
AXE Embed
Nomic-based vector embeddings for Crown search.
Canadian Variants
Gemma3 and Qwen distillations running on edge nodes.

Every Interaction Makes the System Smarter.

↓
User Query
Real-world request
→
↓
Echo Inference
Stream response
→
↓
Logger
JSONL capture
→
↓
SFT
Training corpus
→
↓
Deploy
Updated model
A closed loop. Every query generates training data. Every training run improves the next response. Custody compounds.
25+
Interactions logged today
5K+
Tool-calling filtered dataset
GRPO
Reward shaping for alignment

Own Your Intelligence Layer.

Cloud AI

Models owned by OpenAI / Anthropic / Google
Training data: theirs. Improvements: theirs.
Rate limits. Usage caps. Price increases.
Your data trains their next model.
One API change breaks your product.
No offline capability.

Canadian AI

Models owned by you. Weights on your hardware.
Training data: yours. Improvements: yours.
No limits. No caps. No external pricing.
Your data improves YOUR models only.
You control the API. Forever.
Works perfectly offline.

Performance That Matters.

Model MMLU HumanEval Tool Calling Latency P95
Casanova 2.0 72.4% 68.2% 91.3% 1.2s
Geralt 2.0 64.1% 71.8% 88.7% 0.4s
Anakin 8B 58.3% 63.5% 85.2% 0.18s
Benchmarked on AXE's internal evaluation suite. Real workload, real hardware, real numbers.

What's Next

Month 2
First Canadian fine-tune. Qwen3-8B base + AXE proprietary data. Zero external model dependencies.
Month 3
GRPO reward shaping for tool-calling accuracy. Alignment without cloud compute.
Month 4
Replace Ollama with AXE inference kernel. Native performance. No middleware.
Month 6+
Rust hot-paths for production throughput. Edge model quantization. Hardware acceleration.

Deploy Canadian models
for your organization

/ Contact · we read every inquiry

Talk to AXE.

Demos, partnerships, government RFPs, technical questions. A person reads every form. You hear from someone — not a queue.

Inquiry type

Replies within one business day · Knox audit chain records every inquiry