Models at a glance

The balanced index averages the nine category scores, giving each category 1/9 of the weight regardless of its question count. It measures quality on these test banks, not speed or general model ability. It is not Elo. Only setups with complete, eligible coverage in all nine categories receive a ranked index.

For incomplete results, the range runs from zero to full credit on every unscored task, keeping recorded scores fixed. It is a possible-score range, not a statistical confidence interval. Fully scored but qualified setups show an asterisk and remain unranked. Cells show rounded percentages; select one for exact scores and answers.

By default, all tested thinking modes are included. Each model uses the setup with the most eligible categories, then greatest scored coverage, then its declared baseline. Neither score nor speed chooses it. The same quant, configuration and thinking condition are used throughout a row. Enable all setups to inspect every quant and reasoning effort. Withdrawn setups are omitted.

Size means total parameters, not required RAM. Cloud results describe hosted versions and providers. The two instruction campaigns use identical questions and scorers and share one quality column; evidence IDs remain separate, as do hardware-specific speed comparisons. New, unverified test revisions will not be silently merged. Policy: balanced-categories-v1.