Nemotron 3.5 Super VL is NVIDIA's open-weight AI model, released 15/10/2026. NVIDIA's Hugging Face card describes a 120B-total, 12B-active hybrid Mamba-Transformer mixture-of-experts model with vision inputs, up to 1M tokens of context and the OpenMDW-1.1 licence. Results were collected through NeMo Gym against Qwen3.5-122B-A10B, Qwen3.8-Flash-Next, DeepSeek-V4-Flash0731, Nemotron 3 Ultra, Inkling-Small and GLM-5.3-Flash, and it trails the newer Qwen3.8-Flash-Next on most rows. Recommended sampling is temperature 1.0 and top_p 0.95 for text, 0.6 and 0.95 for vision; the last 11,762 characters of the page, including Ethical Considerations, were not read.
Visit Nemotron 3.5 Super VLAI-Powered
Leverages advanced AI technology to deliver cutting-edge capabilities and results.
Fast & Efficient
Optimized performance ensures quick results without compromising on quality.
Purpose-Built
Specifically designed for llms tasks and workflows.
How does NVIDIA Nemotron 3.5 Super VL score on independent benchmarks?
NVIDIA Nemotron 3.5 Super VL ranks #28 of 28 models on the Artificial Analysis Intelligence Index (v5.0.1) with a score of 13.6, as measured on 22 November 2026.
Where does NVIDIA Nemotron 3.5 Super VL rank against other models?
How does NVIDIA Nemotron 3.5 Super VL do on individual tests?
| Evaluation (run by Artificial Analysis) | NVIDIA Nemotron 3.5 Super VL | Best of any model |
|---|---|---|
| Terminal-Bench 4.0 | 0% | 67.2% |
| FrontierCode 2.0 | 6.1% | 48.8% |
| Humanity's Last Exam | 30.7% | 74.3% |
| GDPval-AA (Elo) | 1162 | 1866 |
How has NVIDIA Nemotron 3.5 Super VL's score changed over time?
We have tracked NVIDIA Nemotron 3.5 Super VL on Intelligence Index v5.0.1 since 22/11/2026. A point is added each day the score, coding index or price changes, so this chart fills in over time.
What does the NVIDIA Nemotron 3.5 Super VL system card report?
NVIDIA's Hugging Face card describes a 120B-total, 12B-active hybrid Mamba-Transformer mixture-of-experts model with vision inputs, up to 1M tokens of context and the OpenMDW-1.1 licence. Results were collected through NeMo Gym against Qwen3.5-122B-A10B, Qwen3.8-Flash-Next, DeepSeek-V4-Flash0731, Nemotron 3 Ultra, Inkling-Small and GLM-5.3-Flash, and it trails the newer Qwen3.8-Flash-Next on most rows. Recommended sampling is temperature 1.0 and top_p 0.95 for text, 0.6 and 0.95 for vision; the last 11,762 characters of the page, including Ethical Considerations, were not read.
These are NVIDIA's own published results (self-reported, with the vendor's harness and settings), so they are not directly comparable with other labs' numbers. Source: nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-BF16 on Hugging Face (15/10/2026). Last checked 22/11/2026.
Knowledge & science
| Benchmark | NVIDIA Nemotron 3.5 Super VL | Qwen3.5-122B-A10B | Qwen3.8-Flash-Next | DeepSeek-V4-Flash0731 |
|---|---|---|---|---|
| GPQA-DiamondCollected via NeMo Gym unless footnoted; BF16 checkpoint | 89.52% | 86.7% | 92.7% | 89.6% |
| HLE (text-only)Collected via NeMo Gym unless footnoted; BF16 checkpoint | 31.79% | 24.2% | 39.6% | 35.8% |
| MMLU-ProXCollected via NeMo Gym unless footnoted; BF16 checkpoint | 76.91% | n/a | 83% | n/a |
| AA-Omniscience accuracyCollected via NeMo Gym unless footnoted; BF16 checkpoint | 26.2% | n/a | 27.9% | n/a |
| AA-Omniscience indexCollected via NeMo Gym unless footnoted; BF16 checkpoint | 4.03 points | n/a | -1.9 | n/a |
Coding
| Benchmark | NVIDIA Nemotron 3.5 Super VL | Qwen3.5-122B-A10B | Qwen3.8-Flash-Next | DeepSeek-V4-Flash0731 |
|---|---|---|---|---|
| LiveCodeBench v6Collected via NeMo Gym unless footnoted; BF16 checkpoint | 88.22% | 76.7% | 94.1% | 94.6% |
| SWE-bench VerifiedCollected via NeMo Gym unless footnoted; BF16 checkpoint | 71.1% | 71.9% | 78.7% | 75.4% |
| SWE-bench ProCollected via NeMo Gym unless footnoted; BF16 checkpoint | 51.78% | 44.9% | 60.37% | 56.91% |
| Terminal-Bench 2.1Collected via NeMo Gym unless footnoted; BF16 checkpoint | 60.67% | 48.7% | 85.8% | 76.8% |
| SciCode (subtask)Collected via NeMo Gym unless footnoted; BF16 checkpoint | 47.13% | n/a | 46.6% | n/a |
| DeepSWE (offline)Collected via NeMo Gym unless footnoted; BF16 checkpoint | 24.65% | n/a | 55.57% | n/a |
Agents & tool use
| Benchmark | NVIDIA Nemotron 3.5 Super VL | Qwen3.5-122B-A10B | Qwen3.8-Flash-Next | DeepSeek-V4-Flash0731 |
|---|---|---|---|---|
| BrowseCompCollected via NeMo Gym unless footnoted; BF16 checkpoint | 78% | 58.1% | 89% | 84.2% |
| tau2-bench telecomCollected via NeMo Gym unless footnoted; BF16 checkpoint | 73.19% | 87.2% | 90.4% | 98.2% |
| tau3-bench bankingCollected via NeMo Gym unless footnoted; BF16 checkpoint | 18.88% | n/a | 44.5% | n/a |
| AutomationBench-AA (offline)Collected via NeMo Gym unless footnoted; BF16 checkpoint | 32.44% | n/a | 72.06% | n/a |
| GDPval-AA v2Collected via NeMo Gym unless footnoted; BF16 checkpoint | 1211 Elo | n/a | 1561 | n/a |
| OSWorld-VerifiedCollected via NeMo Gym unless footnoted; BF16 checkpoint | 51.14% | n/a | 55.8% | n/a |
| WebVoyager (offline)Collected via NeMo Gym unless footnoted; BF16 checkpoint | 77.72% | n/a | 90.39% | n/a |
Safety & honesty
| Benchmark | NVIDIA Nemotron 3.5 Super VL | Qwen3.5-122B-A10B | Qwen3.8-Flash-Next | DeepSeek-V4-Flash0731 |
|---|---|---|---|---|
| AA-Omniscience hallucination rateLower is better; Collected via NeMo Gym unless footnoted; BF16 checkpoint | 30% | n/a | 41.3% | n/a |
Reasoning & maths
| Benchmark | NVIDIA Nemotron 3.5 Super VL | Qwen3.5-122B-A10B | Qwen3.8-Flash-Next | DeepSeek-V4-Flash0731 |
|---|---|---|---|---|
| CritPtCollected via NeMo Gym unless footnoted; BF16 checkpoint | 10.86% | n/a | n/a | n/a |
| APEX shortlistCollected via NeMo Gym unless footnoted; BF16 checkpoint | 88.36% | n/a | 91.8% | n/a |
Long context
| Benchmark | NVIDIA Nemotron 3.5 Super VL | Qwen3.5-122B-A10B | Qwen3.8-Flash-Next | DeepSeek-V4-Flash0731 |
|---|---|---|---|---|
| AA-LCR v1.1Collected via NeMo Gym unless footnoted; BF16 checkpoint | 68.62% | n/a | 78.4% | n/a |
| GraphWalksCollected via NeMo Gym unless footnoted; BF16 checkpoint | 89.71% | n/a | 76.1% | n/a |
| MMLongBench-DocOfficial open-source implementation or internal scaffolding (not yet in NeMo Gym) | 62.82% | 63.7% | 67.9% | n/a |
Other
| Benchmark | NVIDIA Nemotron 3.5 Super VL | Qwen3.5-122B-A10B | Qwen3.8-Flash-Next | DeepSeek-V4-Flash0731 |
|---|---|---|---|---|
| IFBenchCollected via NeMo Gym unless footnoted; BF16 checkpoint | 70.86% | n/a | 75.6% | n/a |
| WMT24++Collected via NeMo Gym unless footnoted; BF16 checkpoint | 84.8% | n/a | 86.2% | n/a |
| LMArena v3 proxyCollected via NeMo Gym unless footnoted; BF16 checkpoint | 52.67 points | n/a | 35.2 | n/a |
| GDP.pdf (offline)Collected via NeMo Gym unless footnoted; BF16 checkpoint; card also lists an Artificial Analysis figure of 6.6 | 6.6% | n/a | 17.4% | n/a |
Multimodal
| Benchmark | NVIDIA Nemotron 3.5 Super VL | Qwen3.5-122B-A10B | Qwen3.8-Flash-Next | DeepSeek-V4-Flash0731 |
|---|---|---|---|---|
| V*Official open-source implementation or internal scaffolding (not yet in NeMo Gym) | 85.6% | 91.1% | 93% | n/a |
| MathVisionOfficial open-source implementation or internal scaffolding (not yet in NeMo Gym) | 78.16% | 83.3% | 87.5% | n/a |
| CharXiv reasoningOfficial open-source implementation or internal scaffolding (not yet in NeMo Gym) | 80.82% | 83.5% | 85.8% | n/a |
| OCRBench v2 (EN)Official open-source implementation or internal scaffolding (not yet in NeMo Gym) | 68.12% | 68.3% | 67.9% | n/a |
| BabyVisionOfficial open-source implementation or internal scaffolding (not yet in NeMo Gym) | 29.6% | n/a | 46.8% | n/a |
| OCR-ReasoningOfficial open-source implementation or internal scaffolding (not yet in NeMo Gym) | 65.08% | n/a | 67.1% | n/a |
| Video-MME v2Official open-source implementation or internal scaffolding (not yet in NeMo Gym) | 32.85% | n/a | 45.1% | n/a |
What did safety testing find for NVIDIA Nemotron 3.5 Super VL?
Safety level: Not published in the readable part of the model card
- Card mentions safety SFT and preference RL datasets in post-training, and says integration needs use-case-specific testing before deployment.
Independent figures: Artificial Analysis, refreshed daily on our AI model benchmarks page. Full analysis: our NVIDIA Nemotron 3.5 Super VL deep dive.
NVIDIA Model Timeline
Up to 1M tokens (some deployment recipes use smaller windows) context
512k tokens context
Specifications
Related Tools
Claude Opus 5.5
Claude Opus 5.5, launched 22 September 2026, is Anthropic's current Opus-line flagship, replacing Opus 5. It performs at roughly the level of Claude Fable 5.1 on most work while costing 40% less to run than Opus 5, at $4 input / $20 output per million tokens (both 20% down, cache reads cut 60% to $0.20). It leads on Terminal-Bench 4.0, GDPval-AA v2.1 and Humanity's Last Exam, and is the first Opus model to ship with Fable-5.1-class cybersecurity, biology and distillation safeguards.
Claude Sonnet 5.5
Claude Sonnet 5.5, released 28 September 2026, is the second model in Anthropic's Claude 5.5 family and the faster, lower-cost complement to Claude Opus 5.5. It keeps Sonnet 5's $2 input / $10 output per million tokens ($0.20 cache reads) while generating output over 30% faster, and scores 70.6% on Terminal-Bench 4.0, above Opus 5.5's 66.4%. It has a 1M-token context window, 128K max output, adaptive thinking, and is the first Sonnet with cyber and anti-distillation classifiers.
GPT-6.1 Sol
GPT-6.1 Sol, launched at OpenAI DevDay on 29 September 2026, is OpenAI's mid-tier workhorse model. OpenAI says it delivers nearly the same intelligence as GPT-6 Astra at one-fifth of Astra's token price: $2 input / $10 output per million tokens with cached input cut to $0.10. It scores 75.2% on DeepSWE v1.1 (high effort) and 71.4% on OSWorld 2.0, has a 1.05M-token context window and 128K max output, and is available in the API plus ChatGPT Work and Codex.
