Nemotron 3.5 Super VL
NVIDIA·
LLMsopen-weightWorkstation
Overview
Vision-language successor to Nemotron 3 Super: a 120B / 12B-active hybrid Mamba-2 + MoE + attention model that reads text, images, documents and GUI screenshots (OpenRouter also lists video input) at a 1M-token context. Shipped under OpenMDW-1.1 with training data and recipes, in BF16, NVFP4 and a released base checkpoint. At Q4 the weights need roughly 70 GB, so it is a single-workstation model rather than a consumer-GPU one. Released alongside the Bolt edge models.
Capabilities and innovations
120B total / 12B active parameters (MoE)Hybrid Mamba-2 + MoE + attention layersText, image, document and GUI-screenshot input1M token context windowReasoning modelBF16, NVFP4 and base checkpointsOpenMDW-1.1 with open training data and recipesFirst vision-language model of the Nemotron 3.5 lineNVFP4 checkpoint with near-BF16 benchmark parity
Benchmarks
| Benchmark | Score | Source |
|---|---|---|
| GPQA Diamond | 89.52% | vendor GPQA Diamond 89.52 BF16 (NVFP4: 89.90). Card not fetchable from this environment (huggingface.co egress-blocked); value taken from consistent secondary transcriptions of the card. |
| Humanity’s Last Exam (no tools) | 30.72% | third party HLE 30.72% no-tools, Artificial Analysis Intelligence Index component (accessed 2026-10-15; Intelligence Index 26.5). NVIDIA's card reports 29.20 full set / 31.79 text-only (BF16). |
| SWE-bench Verified | 71.1% | vendor SWE-bench Verified 71.10 BF16 (NVFP4: 71.07). Era-1 benchmark, recorded only because NVIDIA reports it. Card not fetchable from this environment (huggingface.co egress-blocked); value taken from consistent secondary transcriptions of the card. |
| SWE-bench Pro | 51.78% | vendor SWE-bench Pro 51.78 BF16 (NVFP4: 51.30). Card not fetchable from this environment (huggingface.co egress-blocked); value taken from consistent secondary transcriptions of the card. |
| Terminal-Bench 2.x | 66.7% | vendor Terminal-Bench 2.1 66.70 BF16 (NVFP4: 60.80). Card not fetchable from this environment (huggingface.co egress-blocked); value taken from consistent secondary transcriptions of the card. |
- MMLU-Pro: not reported by the vendor — NVIDIA reports MMLU-ProX (multilingual, 76.91 BF16 / 73.76 NVFP4), a different benchmark; not recorded in the MMLU-Pro field.
- MMMU-Pro: not reported by the vendor — No MMMU-Pro figure in the model card as transcribed; Artificial Analysis has not run it either.
- Terminal-Bench 4.0: not reported by the vendor — NVIDIA lists 'Terminal Bench 4 (offline)' at 1 (BF16) / 1.5 (NVFP4) — an offline variant, not the Harbor-harness 4.0 run the field holds. Artificial Analysis scores 0 on its own Terminal-Bench 4.0 run, a harness on which most frontier models also score 0, so neither figure is comparable with the vendor 4.0 values in the catalog.
- DeepSWE v1.1: not reported by the vendor — NVIDIA lists 'DeepSWE (offline)' at 24.65 BF16 / 18.2 NVFP4 without a version label and outside the fixed mini-swe-agent harness; only explicitly labelled v1.1 values go in the field.
Architecture and hardware
- Parameters
- 120B total, 12B active per token (MoE)
- Estimated VRAM at Q4
- ~69 GB, Workstation class
- Quantization formats
- BF16, NVFP4
- Recommended runtime
- vLLM
- License
- OpenMDW-1.1
Links
More from NVIDIA
Data curated by AI Model Timeline. See the methodology for admission criteria, benchmark eras and source priority.