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.

Flagged models are greyed out by default. Use the checkbox to hide them from results, charts, timeline and headline stats. Flags identify disclosed benchmark-specific training or unusually strong scores against comparable-size peers. These are review signals, not proof that results are false or that test data leaked. Each flagged model includes the reason and a source.

How score outliers are flagged

Compare other organizations’ models within half to twice the parameter count, with at least 5 peer results from 3 organizations. BLiMP / ARC-Easy must exceed the peer median by more than both 10 percentage points and 3 robust standard deviations (1.4826 × median absolute deviation). WikiText-2 must be more than 3× better and 3 robust standard deviations below the median in log space. ACI is excluded because it uses a different measurement. Out-of-range scores are also flagged. Peer reports are unverified; different training budgets and evaluation protocols can explain outliers.

Detailed Results

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