ResearchBenchmarksCode Generation on HumanEval (Accuracy and Loss)Follow63AccuracyNexus3743.7550.557.25Apr 10, 2026Evaluation ResultsMethodMethodLinksAccuracyLossNexusModel scale=3BModel scale=3B2026.04631.199AdamModel scale=3BModel scale=3B2026.04591.205NexusModel scale=1BModel scale=1B2026.04411.325AdamModel scale=1BModel scale=1B2026.04381.324