Lumina indices / General
Factual reliability, instruction following and document understanding under defined protocols.
Lumina-Core-v1-4c99ab21f8b59ce1bd75
Public projection cutoff: 2026-09-13. The cutoff identifies a retained source snapshot, not universal freshness. Complete baselines have no qualified broad interval. Qualified GI prediction ranges remain visible for their specific missing-component targets.
Models with category evidence
293
Exact model and checkpoint identities.
56 scored in Default · 56 with 2025+ release dates.
Distinct exact canonical checkpoints with reviewed native category evidence in the retained Core corpus or a reviewed category assessment. Includes unscored and historical releases; excludes configuration duplicates, explicitly routed identities and merged aliases. Open Weights is restricted to its verified parent cohort.
Benchmarks represented
3
Distinct definitions in the retained evidence.
3 featured views selected for breadth and comparable records.
Distinct benchmark registry definitions represented by qualified native category evidence. Protocols, source mirrors, task children, configurations and conceptual families are not added as separate benchmarks. LiveBench tasks share one definition; featured panels are a nonredundant selection. Auxiliary tracks without a matching reviewed registry definition are excluded from this count.
Lumina General Index
Complete baselines have no qualified broad interval. Qualified GI prediction ranges remain visible for their specific missing-component targets.
Chart loads as you explore
Lumina General Index
42 of 56 default 2025+ releases have a compatible standard API price.
Chart loads as you explore
Lumina General Index
43 of 56 default 2025+ releases have a compatible developer-API throughput measurement.
Chart loads as you explore
Original measurements
Reviewed source records, with each task’s own units and execution conditions.
Artificial Analysis · Signed utility · percentage points · Higher is better
LiveBench · % · Higher is better
Artificial Analysis · % · Higher is better
3 distinct featured views. Native benchmark positions are separate from index ranks. Imported measurements are not independent replications; the complete tables retain source and configuration details.
From evidence to insight
Real tasks, documented sources and defined execution conditions.
Factual reliability, instruction following and document understanding under defined protocols.
An equal-third native composite of factual utility, instruction following and text-document understanding. Signed factual utility rewards correct answers and penalizes incorrect ones; it is not simple accuracy. Supplemental source inputs remain non-scoring. Chart positions retain full precision; display rounding does not change their order.
A point lead does not establish superiority. Provisional estimates stay visibly marked.
43 points have no qualified range. Their scores remain available; missing ranges are not zero uncertainty.
Complete baselines have no qualified broad interval. Qualified GI prediction ranges remain visible for their specific missing-component targets. Absence from Default is not evidence of inferiority.
