TibiOS
Presenting
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 โ all within the same infrastructure? That is what we achieved with TibiOS. Zero token spend.
SAURON is based on Qwen3-4B, adapted with LoRA on a heterogeneous fleet of five machines:
16 GB โ training with MLX.
16 GB โ training with MLX.
8 GB each โ training with Unsloth + TRL on CUDA.
16 GB โ head node and router.
The inference lanes generate their own training data.
The Macs train in a ring, splitting each batch by every machine's measured speed.
Training produces the SAURON LoRA adapter.
The head node publishes the adapter as a new inference lane: @tuned.
The fleet infers again โ and can generate new data for the next cycle.
Base model, quantized.
MLX on Apple Silicon; Unsloth + TRL on CUDA.
Rank 8, alpha 16, seven projections.
Trainable โ about 0.74% of the model.
Adapter size.
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 โ not to train SAURON.
The important result is not just a tuned model. It is a working cycle in which the same fleet both trains and infers โ 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.