Nano Chat

An open-source project implementing the complete LLM pipeline: tokenization, pre-training, fine-tuning for chat, evaluation, and a web UI for interacting with the result.

What It Does

Lets you train your own small language model from scratch for approximately $100 in GPU time. The result won’t compete with frontier models (Claude, GPT, Gemini), but gives you a model under your absolute control — custom training data, custom behavior, no external dependencies.

Why It Matters

Democratizes LLM training to the point where an individual developer can afford to experiment. The practical use cases are niche (custom domain models, research, education about how LLMs work), but the symbolic value is significant: for $100, you can understand and own the entire stack.

See Also

• llama.cpp — for running pre-trained models locally (vs training your own)
• AutoResearch and Evals — methodology for iterating on model/skill quality
• Fireship — source
• Source: 7 Open-Source AI Tools