The archive · Product Ideas · Product decision · 2026
GuppyLM: an 8.7M-parameter fish LLM anyone can train in five minutes
An 8.7M-parameter fish-voiced LLM trained from scratch in one Colab notebook and five minutes — demystifying the whole pipeline.
GuppyLM
What it had to solve
The developer behind GuppyLM wanted to show that training a language model is not magic — no PhD, no GPU cluster required — so the model had to be small enough that data generation, tokenizer, architecture, training loop and inference could each be seen and rerun, and cheap enough that a free Colab session and five minutes would do.
How it works
Most people meet language models only as opaque, billion-parameter services. GuppyLM, built by the developer behind GitHub's arman-bd, exists to show that training your own LLM is not magic: no PhD, no GPU cluster, one Colab notebook, about five minutes, and you have a working model you built from scratch — data generation, tokenizer, architecture, training loop and inference. The model plays a small fish named Guppy who lives in a tank, talks about water, food, light and bubbles, and cheerfully admits it does not understand money, phones or politics.
Every part of the project is deliberately small and transparent. The dataset is 60,000 synthetic conversations across 60 fish-life topics, generated by template composition with randomized components — 30 tank objects, 17 food types, 25 activities — so one consistent personality stays believable. The model is a vanilla transformer with 8.7 million parameters: six layers, 384 hidden units, six heads, a 4,096-token BPE vocabulary and a 128-token context, with no GQA, no RoPE, no SwiGLU and no early exit.
The design decisions are documented as tradeoffs rather than hidden. Personality is baked into the weights because a 9M model cannot conditionally follow a system prompt; training is single-turn because a fish that forgets is on-brand while garbled output is not; and a T4 GPU with five minutes is enough at this size. The project exports a quantized ~10 MB ONNX model that runs entirely in the browser through WebAssembly — no server, no API keys — and a Colab notebook trains, tests and uploads your own Guppy.
The Show HN of 6 April 2026 drew 915 points and 134 comments, and the repository showed 3.4k stars and 303 forks when the material was collected in September 2026. GuppyLM does not try to write essays; its wager is that a tiny, honest, characterful model does more to demystify the pipeline than any diagram, because the reader can rebuild the whole thing in an afternoon.
Why it lands
- The pipeline is end-to-end and tiny: dataset generation, tokenizer, transformer, training and inference all live in one notebook a single developer can finish in an afternoon.
- Character does real pedagogical work: the fish persona's limits are on-brand, so the small context and single-turn design read as charm rather than failure.
- Constraints are stated as deliberate tradeoffs with reasons — no system prompt, single-turn only, vanilla attention — which teaches judgment instead of recipes.
- The result is portable proof: a quantized ~10 MB ONNX model runs in the browser via WebAssembly, so anyone can chat with a model they could have trained themselves.
- The reception — 915 HN points, 134 comments, 3.4k stars — showed the audience for 'small enough to rebuild' is real.
What it did
The Show HN of 6 April 2026 drew 915 points and 134 comments; the repository showed 3.4k stars and 303 forks when the material was collected in September 2026 — evidence that a tiny, honest, characterful model made the training pipeline feel graspable rather than magical.
What you can take
To demystify a technology, shrink it until one person can rebuild it in an afternoon and each simplification is a stated tradeoff — then give the artifact a personality so its limits read as charm.
Since then
GuppyLM stayed MIT-licensed and open source (24 commits at collection): a browser demo running the quantized ONNX model through WebAssembly, a Colab notebook that trains a fresh fish on a T4 GPU, and a HuggingFace dataset holding the 60K synthetic conversations. The Show HN of 6 April 2026 drew 915 points and 134 comments; by September 2026 the repo had 3.4k stars and 303 forks. The wager stays what it was at launch: a nine-million-parameter fish you can train in five minutes demystifies models more than any diagram, because the reader can rebuild the whole pipeline in an afternoon.
Sources
spotted an error? The archive wants to know.
Your turn
You just read one. Describe the brief you are staring at, and see who has been given the same problem.
Free account · 3 free questions · no card