I’m 15 years old and I build small language models from scratch: tokenizer, pre-training, instruction fine-tuning and evaluation.

I have already trained and published G0 Nano and G1 Nano, two open models of around 60 million parameters, using only a Jetson Orin Nano with 8 GB of memory.

My latest project is G2 Nano Safety, a specialized 60M-parameter model designed to filter prompts before they are sent to an AI agent. It can let a prompt through, remove sensitive information, or block it. Despite being much smaller than Shieldstral, Mistral AI’s 3B safety model, it performs better on this specific task. The model is still evolving, so I am not publishing detailed benchmark numbers yet.

The Jetson is now becoming a serious limitation. I’m raising funds to acquire an NVIDIA DGX Spark, with 128 GB of unified memory — 16 times more memory than my current machine. This would allow me to train larger and more capable open models locally, and to document the entire process publicly.

If the fundraising goal is not reached, every contribution will still be used for the project: cloud GPU credits, storage and training costs.

I will document the progress, expenses, experiments, successes and failures openly.