Smol Intelligence Index

An independent benchmark dedicated to evaluating sub-400M parameter language models. We measure foundational reasoning, commonsense, and arithmetic capabilities to track the frontier of small-scale AI.

Leaderboard

Standardized zero-shot evaluation. Higher scores are better. Click any column header to sort, and hit the + icon to build a custom comparison.

Size:
+ # Model Params Smol Int Index HellaSwag ARC-Easy ARC-Challenge PIQA ArithMark-3 ArithMark-2

Side-by-Side Comparison

Direct comparison of your selected models. Top scores within this selection are highlighted.

× Model Params Smol Int Index HellaSwag ARC-Easy ARC-Challenge PIQA ArithMark-3 ArithMark-2

Efficiency Rankings

Intelligence Index per Million Parameters. This metric reveals which architectures deliver the most cognitive power relative to their size.

# Model Params Smol Int Index Efficiency (Idx / M)

Rules

Eligibility, evaluation, and submission requirements.

RULE 01 The 400M Parameter Limit. This board is strictly for small-scale models. To qualify, your architecture must have fewer than 400 million parameters.
RULE 02 Original Pretraining Only. We only accept models trained entirely from scratch by the submitting individual or organization. No fine-tunes or merged weights allowed.
RULE 03 Fully Open Weights. Your model's weights must be publicly accessible. We require the actual training checkpoint, not a post-processed or merged variant.
RULE 04 Mandatory Verification. Trust is everything. Before any new entry is published, our team will manually audit and replicate your benchmark scores to ensure total accuracy and fairness.
RULE 05 How to Submit. Ready to add your model? Simply open a discussion or pull request on our Hugging Face Space with your evaluation logs and model links.

Add your model

To get listed, open a discussion on our Space containing your evaluation logs and repository links. Remember, your model must be trained from scratch with fully open, unmerged checkpoint weights to pass our verification process.

How the Custom Int Index is calculated

Our index applies specific weights to each benchmark component to balance different cognitive domains. Performance falling below random chance is clamped at zero.

Custom Int Index =

[ (100 × max(0, (ARC-Easy - 25) / 75) × 0.75) +
  (100 × max(0, (ARC-Challenge - 25) / 75) × 1.20) +
  (100 × max(0, (HellaSwag - 25) / 75) × 0.9) +
  (100 × max(0, (PIQA - 50) / 50) × 0.75) +
  (ArithMark3 × 0.75) +
  (ArithMark2 × 0.40) ] / 4.75

Weights: ARC-Easy weight 0.75 (chance 25), ARC-Challenge weight 1.20 (chance 25), HellaSwag weight 0.90 (chance 25), PIQA weight 0.75 (chance 50), ArithMark-3 weight 0.75, ArithMark-2 weight 0.40.