Priming Base Models for Agentic Post-Training: An Open Midtraining Recipe

Open paper
Benchmark Model Rank Results
code-generation-on-livecodebench-v6NVIDIA Nemotron 3.5 Super VL 120B-A12B BF16–Pass@1: 0.8822
code-generation-on-scicodeNVIDIA Nemotron 3.5 Super VL 120B-A12B BF16–Accuracy: 47.13
code-generation-on-swe-bench-multilingualNVIDIA Nemotron 3.5 Super VL 120B-A12B BF16–Accuracy: 67.72
code-generation-on-swe-bench-verifiedNVIDIA Nemotron 3.5 Super VL 120B-A12B BF16–Accuracy: 71.1
code-generation-on-terminal-bench-2-1NVIDIA Nemotron 3.5 Super VL 120B-A12B BF16–Accuracy: 60.67
computer-use-agents-on-osworld-verifiedNVIDIA Nemotron 3.5 Super VL 120B-A12B BF16–Success Rate: 51.14
document-question-answering-on-mmlongbench-docNVIDIA Nemotron 3.5 Super VL 120B-A12B BF16–Accuracy: 62.82
multimodal-reasoning-on-mathvisionNVIDIA Nemotron 3.5 Super VL 120B-A12B BF16–Accuracy: 78.16
optical-character-recognition-ocr-on-ocrbench-v2NVIDIA Nemotron 3.5 Super VL 120B-A12B BF16–English Score: 68.12Chinese Score: 59.15

Rank counts only results with a code link.