Model Training

Train small models in your browser, or use your own GPU for advanced fine-tuning. You own everything — the data, the weights, the model.

The training pipeline
01·Prepare data

Export your text as a .txt file. 50K+ characters recommended for good results. Books, articles, or personal writing all work.

02·Train with LoRA

SmolLM2-360M is loaded in 4-bit via Unsloth. LoRA adapters (rank 16) are trained on your text. Takes 5-15 min on a T4 GPU.

03·Export to ONNX

The adapted model is merged and exported to ONNX with full KV-cache support. Zipped for browser upload.

04·Chat in browser

Upload the ZIP to the Chat page. All inference runs locally via WebGPU/WASM — no server needed.

Training Methods

Personality

Training in action

Set up your workspace, prepare your data, and train your custom AI personality

Zero-cost philosophy

Training runs on your hardware. Browser training uses your CPU. Advanced fine-tuning uses your GPU, Google Colab, or any machine you choose. We provide the platform — your computer does the work. No API fees, no cloud costs, no limits.