Benchmarks
Benchmarks

Arena Elo

Human preferenceunit Elo · normalized over [1000, 1480]

Blind pairwise human votes (arena.ai). The gold standard for conversational preference.

#ModelSourceScoreNormalized
Score distribution
100 tracked results across the normalization window
10001480
Score vs. parameters
Open-weights models, log-x params
1B10B100B1000B