REDDIT · REDDIT// 4d agoMODEL RELEASE

MeowLLM builds tiny cat persona from scratch

MeowLLM is a ~3.5M-parameter language model trained from scratch to speak as Miso, a house cat, with the personality encoded in the weights rather than via prompting or fine-tuning. The repo includes the full pipeline: data generation, tokenizer training, masking, evaluation, inference, documentation, and tests, plus a Colab path to chat with the pretrained checkpoint or train your own. It is positioned as a compact, educational end-to-end example of building a character model rather than a practical general-purpose LLM.

// ANALYSIS

Fun demo, solid engineering exercise, limited real-world utility.

–The interesting part is not raw capability; it is the claim that a character voice can emerge from training data and architecture alone.
–The repo looks unusually complete for a tiny model project: data gen, eval harness, tests, and inference are all part of the package.
–At 3.5M parameters, this is firmly in novelty/education territory, not a competitive assistant model.
–The cat persona angle is memorable and likely helps it spread, which is probably the point.
–Best fit is as a teaching artifact for small-model training workflows and character modeling, not as a deployable product.

// TAGS

llmtiny-modelcharacter-modelfrom-scratchopensourcepytorchlocal-llmcatcolab

DISCOVERED

4d ago

2026-04-07

PUBLISHED

5d ago

2026-04-07

RELEVANCE

8/ 10

AUTHOR

Dismal_Beginning_486