AXE Specialists
Collection
Fine-tuned models, each purpose-built for a specific lane in the inference pipeline. • 7 items • Updated
Caesar 1.0
Our first complete training cycle, end to end. SFT followed by GRPO policy optimization. ~1B parameters of weights we own.
Caesar isn't a fine-tune of someone else's model. It's a full training run we executed on our own pipeline — supervised fine-tuning on tool-calling data, then GRPO over 600+ iterations to optimize the policy. The point isn't that 1B beats the frontier on benchmarks. It's that we control every byte of how it got there.
Model details
| Property | Value |
|---|---|
| Developer | AXE Technologies |
| Training | Custom SFT + GRPO (600+ iterations) |
| Parameters | ~1B |
| Context | 8K tokens |
| Quantization | GGUF |
| License | Apache 2.0 |
What it's tuned for
- Fast edge inference on minimal hardware
- Tool calling and function dispatch
- Offline operation — no cloud dependency for any inference step
- Demonstrating end-to-end training capability on Apple Silicon
Usage
ollama pull axetechnologies/caesar-1.0
The AXE family
Five models, each tuned for a different lane in the inference pipeline:
| Model | Lane | What it does |
|---|---|---|
| Casanova 1.2 | Agency | Tool-calling, multi-step workflows. 27B dense. |
| Geralt 1.3 | Reasoning at scale | 26B parameters of capability, 4B of inference cost. MoE. |
| Pegasus 1.0 | Visible work | Chain-of-thought you can audit. 12B dense. |
| Artemis 1.0 | Speed | Loads in seconds. 4B for edge hardware. |
| Caesar 1.0 | First principles | Our own training cycle. ~1B, end-to-end on our pipeline. |
About AXE Technologies
Canadian in-house AI infrastructure. Built on Apple Silicon. The models run on hardware you can audit — no cloud dependency, no third-party model in the data path.
Website: axetechnologies.ca