nanochat playground

Compare four published nanochat checkpoints from a single prompt. SFT/RL models reply as a chat assistant; base models autocomplete the prompt.

Built for MIOTI Master MDL2603A — IA Generativa, Sesión 2.
Model

Larger models are slower but more capable. First load downloads weights from HF.
16 512
0.1 1.5
1 200

Notes

  • base models are pre-trained only — they continue your prompt rather than answer it.
  • SFT/RL models follow chat-style prompts; output stops at <|assistant_end|>.
  • First request after model switch will download checkpoint files (~0.6 - 2.2 GB) and may take a couple of minutes.