Nemotron 3.5 Super VL

Changes from previous version

Adds vision and a 120B-total, 12B-active backbone after the compact text model Nemotron 3.5 Lightning. Context stays at 1,048,576 tokens. NeMo Gym Terminal-Bench 2.1 is 60.67. Median hosted price is $0.20/$0.575 per million tokens, and measured price per task is $0.25.

Release Summary

NVIDIA's open vision-language model for always-on agents. 120B total parameters, 12B active, a hybrid Mamba-2 and mixture-of-experts backbone with a C-RADIO vision encoder, and OpenMDW-1.1 weights in BF16 and NVFP4. Text, image, and video in. The prompt and the answer share a 1,048,576-token context, and NVIDIA does not publish a separate output cap. On the model card's NeMo Gym table, SWE-bench Verified is 71.10, SWE-bench Pro is 51.78, and Terminal-Bench 2.1 is 60.67. CharXiv reasoning is 80.8%, against 63% for Nemotron 3 Nano Omni. Hermes Bench is 63.44%, against 56.33% for Nemotron 3 Ultra. Median hosted price is $0.20/$0.575 per million tokens. Measured price per task is $0.25. NVIDIA reports up to 5x higher throughput than other open models in this size class. Same-day edge model: Nemotron 3.5 Bolt.

Timeline

Other Nemotron versions