🧬 Genesis 1B — Live Training Playground

A 1 billion parameter language model being pre-trained from scratch on 2× RTX 4090s. No cloud compute. No corporate backing. One person, one workstation.

Built by Kroonen AI — read the technical blog post for the full story.

This is a raw base model — it completes text, it doesn't chat. Select a checkpoint and watch the model evolve from random noise to coherent language as training progresses. Earlier checkpoints = more noise. Later checkpoints = emergent structure and fluency.

Training checkpoint
10 500
0.1 2
0 200
0.1 1

Genesis is being trained live — new checkpoints appear as training progresses. Hit 🔄 Refresh to check for new ones.
Inference runs on HuggingFace ZeroGPU (NVIDIA H200). Base model weights are currently private while the model undergoes alignment training (SFT/DPO). The final aligned model will be released under Apache 2.0, with continued training producing improved revisions over time.