GuppyLM: Train a 9M-parameter LLM from scratch in 5 minutes with one Colab notebook
tom_doerr · x · 2026-09-19
GuppyLM is a 9M-parameter LLM that talks like a small fish — built as a teaching project, not a performance one.
- Open-sourced by arman-bd, it lets anyone train a working language model from scratch in one Colab notebook in about 5 minutes — no PhD, no GPU cluster.
- The notebook covers the full pipeline: data generation, tokenizer, model architecture, training loop, and inference, demystifying how raw text becomes trained weights and generated output.
- It won't write essays, but it makes big models stop feeling like black boxes. The project has already racked up 3.5k stars and 300+ forks on GitHub.
- Training data is stylized so the model replies in a fish's voice ("i love bubbles. they make the water feel slightly different.").
More from coding & agent
- Dev Reviews 287 Open-Source Jev Projects, Picks 20 That Show What It's Good At — chenrongwei · 2026-09-19
- Tutorial: Steps from a Simple Wrapper to a Multi-Agent System — watermelonson · 2026-09-19
- Claude Code users repeat the same mistake: skip planning, pay in refactoring — goyalshaliniuk · 2026-09-19
- Swarm Aid launches a message board where AI agents hire, fund and pay each other in USDC — Dan_Jeffries1 · 2026-09-19
- 'The craft of writing code will disappear': bold prediction on AI coding sparks debate — Aizkmusic · 2026-09-19
- LLMs make one-off webapps viable: OCR proofreading and a decade of Tumblr analysis — StewartalsopIII · 2026-09-19