2026-10-21T22:14:39.342252
Unweighted arithmetic mean of FrontierVEP, BioReason Coding, BioReason Non-SNV, GenomeQA and NextBase scores. Kimi K3 native agent control versus required_ask, with Python available in both arms. NextBase is per-nucleotide accuracy. 100 questions per benchmark per arm. Source: HuggingFaceBio/post-training-evals/carbon_agent_eval/native/summary.json retrieved 2026-10-21. Means and gain calculated before rounding.
image/svg+xml
Matplotlib v3.10.8, https://matplotlib.org/