Four H100s or two GB200s. That is NVIDIA's reference footprint for serving Nemotron 3.5 Super VL 120B A12B in full BF16 precision, and it tells you who this model is for: teams that want a frontier-class open reasoning model they can host themselves, with vision and video built in rather than bolted on.
Released on 15 October 2026 under the OpenMDW-1.1 licence, it succeeds the text-only Nemotron 3 Super 120B A12B.
Architecture in brief
It is a hybrid Mamba-2 and Transformer mixture-of-experts with 120B total and 12B active parameters, 512 routed experts per MoE layer, multi-token prediction, and a C-RADIO vision encoder. Pretraining covered about 26 trillion tokens, and NVIDIA released open datasets and recipes alongside the weights.
The multi-token prediction head doubles as a built-in drafter for speculative decoding, so no separate draft model is needed.
Spec sheet
| Item |
Detail |
| Model ID |
nvidia/nemotron-3.5-super-vl-120b-a12b |
| Weights |
nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-BF16; NVFP4 build also published |
| Parameters |
120B total / 12B active |
| Context |
Up to 1M tokens |
| Input |
Text, JPEG/PNG images, MP4 video up to 2 minutes |
| Output |
Text |
| Reasoning |
Off, medium effort, or full |
| Licence |
OpenMDW-1.1 |
| Indicative hosted price |
$0.15 in / $0.50 out per 1M tokens (OpenRouter); free tier offered |
Thinking, three ways
Reasoning is controlled from the chat template: enable_thinking true or false, and medium_effort for a middle setting that spends fewer tokens at some cost to accuracy. Reasoning and final-answer tokens share the max_tokens budget, so set it generously when thinking is on.
Publisher-reported scores
NVIDIA reports GPQA-Diamond 89.5, SWE-bench Verified 71.1, LiveCodeBench v6 88.2, MathVision 78.2 and OSWorld-Verified 51.1. Treat them as a shortlist reason, not a verdict.
Inconsistencies to resolve yourself
- Languages. NVIDIA's metadata says six languages; its training notes cite English plus 19 others. Test the languages you need.
- Max output. One listing gives 8,192 output tokens; OpenRouter allows a maximum completion of 235,929. Check the limit on whichever endpoint you call.
- Sampling. The summary recommends temperature 0.6 for vision; the vision examples use 1.0. Try both.
- A "text-only" listing. Some catalogues carry a text-only
nemotron-3.5-super-120b-a12b. NVIDIA publishes only the VL checkpoints, so treat that as this model.
Sizing a pilot
Start on a hosted endpoint (NIM on build.nvidia.com, OpenRouter, or ModelScope) before buying GPUs. Run your agent traces with thinking off, medium and full, and log tokens per task as well as success rate; medium often wins on cost. If you self-host, vLLM and SGLang have documented recipes, and the NVFP4 build targets Blackwell. Compare against Qwen3.5 122B-A10B, a similar-sized open MoE.
More from NVIDIA on its provider page, and peers under image-text-to-text. Hosted prices are indicative; confirm on your account.