Nemotron 3.5 Super VL

// Overview

Nemotron 3.5 Super VL is the larger of the two models NVIDIA added to its Nemotron 3.5 generation on 15 October 2026, published alongside the laptop-sized Nemotron 3.5 Bolt 74B-A9B. It is a single open model that handles reasoning, instruction following and tool use together with image, document and video understanding, so one checkpoint can sit behind an agent that reads charts, scanned forms and screen recordings as well as text. NVIDIA built it by distilling from its larger Nemotron 3 Ultra and describes its accuracy as comparable with that model at roughly one-fifth of the parameter count.

The architecture follows the rest of the Nemotron 3 and 3.5 families: Mamba-2 state-space layers interleaved with attention layers over a Mixture-of-Experts stack, here with 120 billion total parameters, 12 billion active per token and 512 routed experts per MoE layer. A C-RADIO vision encoder and a vision-language connector feed images (JPEG, PNG) and MP4 video up to two minutes long into the language model, and built-in Multi-Token Prediction supports speculative decoding. The context window runs to one million tokens. NVIDIA's announcement puts training at about 26 trillion tokens across pre-training and continued pre-training, followed by roughly 3.8 million post-training samples and reinforcement learning across dozens of environments, including 10 billion tokens of continued pre-training on Indic languages.

As with Lightning, the release is open in the full sense NVIDIA uses for Nemotron: base, post-trained and NVFP4-quantised checkpoints under the OpenMDW-1.1 license, plus training recipes, datasets (including 40 vision-language datasets) and RL environments. The card lists 4× H100 or 2× GB200 for the BF16 checkpoint. Weights are on Hugging Face and ModelScope, hosted access runs through NVIDIA NIM on build.nvidia.com and OpenRouter, and NVIDIA names vLLM, SGLang, Ollama, LM Studio and Unsloth among supported runtimes.

NVIDIA's own comparison table is candid. Super VL scores 89.52 on GPQA Diamond, 71.10 on SWE-bench Verified, 60.67 on Terminal-Bench 2.1 and 80.82 on CharXiv reasoning, which puts it ahead of Qwen3.5-122B-A10B on most rows but behind Qwen3.8-Flash-Next and GLM-5.3-Flash on most coding and agent benchmarks — Terminal-Bench 2.1 is 85.80 for Qwen3.8-Flash-Next and 82.70 for GLM-5.3-Flash. NVIDIA's case is efficiency: it reports up to 5× higher agentic throughput on a single GPU than other open models in its size class and over 600 output tokens per second on Artificial Analysis agentic tasks on a Baseten endpoint.

// Benchmarks

This model's scores

1. GPQA Diamond89.52%
2. LiveCodeBench v688.22%
3. CharXiv (reasoning)80.82%
4. MathVision78.16%
5. BrowseComp78%
6. SWE-bench Verified71.1%
7. IFBench70.86%
8. AA-LCR v1.168.62%
9. OCRBench v2 (English)68.12%
10. Terminal-Bench 2.160.67%
11. SWE-bench Pro51.78%
12. OSWorld-Verified51.14%
13. Humanity's Last Exam (text-only)31.79%

Scores on a 0–100 scale (25-point gridlines); higher is better. Each benchmark links to its published source.

// Strengths

• One open checkpoint covers text reasoning, tool use, OCR, charts, documents and up to two minutes of video
• 12B active parameters out of 120B, with NVIDIA reporting up to 5× the single-GPU agentic throughput of open models in its class
• Permissive OpenMDW-1.1 license with base, post-trained and NVFP4 checkpoints, training data, recipes and RL environments published
• 1M-token context window and native Multi-Token Prediction for speculative decoding
• 80.82 on CharXiv reasoning and 89.52 on GPQA Diamond in NVIDIA's published evaluations

// Best for

• Document-heavy agents — invoices, receipts, contracts, charts and tables — where text and vision need to live in one model
• Self-hosted multimodal agents on DGX, RTX PRO or cloud GPUs where a permissive license and published training data matter
• Post-training and domain customisation from the open BF16 and base checkpoints with NeMo AutoModel, NeMo RL and NeMo Gym
• Indic-language workloads, drawing on the 10B tokens of Indic continued pre-training NVIDIA describes

// How to access

// Nemotron (open-weight) — every version

The full lineage of the Nemotron (open-weight) line, newest first. Every version has its own page — click any to compare specs, benchmarks and pricing.