catmind-1.2b: A meme model that trades reasoning for cat stories

The AI community has a new joke model: catmind-1.2b, a fine-tune of LFM2.5-1.2B-thinking that replaces any form of rational reasoning with cat stories. Created by Reddit user marcodsn, the model uses its thinking block to narrate cat anecdotes completely unrelated to the user's query. Benchmarked on the Crucible dataset, catmind-1.2b achieved a mere 24.3% accuracy, far below the base thinking model's 75.6% and even below the instruct-only version's 49.2%. The model also outputs fewer tokens on average (1,194) compared to its reasoning counterpart (4,243).

The creator tested whether the model might still reason in hidden states while outputting cat stories, but found no evidence of that — accuracy remained the same even when pre-filling a story. marcodsn calls it a “meme model” and explicitly warns against using it for any serious work. Still, it's an amusing exploration of what happens when you intentionally cripple a reasoning model. The release underscores how critical the thinking block is for performance — and provides a lighthearted reminder that not every fine-tune is a step forward.