Presenting

SAURON

A model trained by its own fleet.

What if a set of home computers could generate their own training data, train a model, deploy it and use it for inference — all within the same infrastructure? That is what we achieved with TibiOS. Zero token spend.

The fleet that trained it

SAURON is based on Qwen3-4B, adapted with LoRA on a heterogeneous fleet of five machines:

🍎 Mac mini M4

16 GB — training with MLX.

💻 MacBook Air M1

16 GB — training with MLX.

⚡ 2 × Jetson Orin Nano

8 GB each — training with Unsloth + TRL on CUDA.

🔌 Orange Pi 6+

16 GB — head node and router.

The full pipeline

1

The inference lanes generate their own training data.

2

The Macs train in a ring, splitting each batch by every machine's measured speed.

3

Training produces the SAURON LoRA adapter.

4

The head node publishes the adapter as a new inference lane: @tuned.

5

The fleet infers again — and can generate new data for the next cycle.

Technical data

Qwen3-4B · 4-bit

Base model, quantized.

MLX + Unsloth/TRL

MLX on Apple Silicon; Unsloth + TRL on CUDA.

LoRA r=8, α=16

Rank 8, alpha 16, seven projections.

16.5M parameters

Trainable — about 0.74% of the model.

64 MB

Adapter size.

Home Wi-Fi LAN

Zero electricity cost in our scenario, no cloud needed for this local pipeline.

Kimi K3 and Claude were used exclusively to draft the glossary data — not to train SAURON.

SAURON infographic: the fleet, the pipeline and the LoRA adapter data

The important result is not just a tuned model. It is a working cycle in which the same fleet both trains and infers — a fundamental piece of the TibiOS vision: distributed AI infrastructure where heterogeneous machines collaborate, learn and serve models.

You don't need to start with a data center.
Start with what you have.