Tiny-ML Leaderboard

Sub-150M parameter language models, same eval harness, transparent methodology.

Every tiny LM with verifiable benchmarks — ours, our competitors', yours. Submit a model via PR.

Scoring. Efficiency (⚡) is the overall benchmark score (average of BLiMP, ARC-Easy and normalised WikiText-2) times a size bonus, capped on a log scale: the smallest model on the board tops out at 1.5×, the largest gets 1.0×. A tiny model can't out-rank a much larger, better-performing one on parameter count alone.
Unknown org? That tag is for makers submitting benchmarks ahead of release — DM glintresearch on Discord to have it replaced with your org.

Detailed Results

Sort

#	Model	Org	Params	Eff. ⚡	WikiText-2 byte_ppl ↓	BLiMP ↑	ARC-Easy ↑	GCI ↑	Training Tokens	Released	Links

Tiny-ML Leaderboard

Detailed Results

Model Release Timeline

Benchmark Overview

BLiMP

Model Efficiency

Parameters vs Leaderboard Score