NVIDIA Releases Nemotron 3.5 Super VL for Agentic Multimodal Workflows
NVIDIA launches an open 120B mixture-of-experts model alongside compact edge variants.
1Open-Weight Architecture Launch
NVIDIA released Nemotron 3.5 Super VL on October 15, 2026, as an open, customizable multimodal model. It utilizes a 120-billion-parameter mixture-of-experts architecture that activates 12 billion parameters per token. The system targets specialized, always-on enterprise agents handling multistep workflows.
2Unifying Multimodal Pipelines
The model consolidates reasoning, instruction following, and tool use with text, image, document, and video understanding. This unified design removes the operational requirement of running separate vision-language models alongside text models, eliminating the latency and context serialization overhead of passing data between disparate systems.
3Throughput and Hardware Efficiency
On a single GPU, Nemotron 3.5 Super VL delivers up to 5x higher throughput for agentic tasks compared to peer open models. Benchmark testing by Artificial Analysis placed it ahead of competitors like InclusionAI's Ling 3.0 Flash and DeepSeek's V4.1 Flash on task completion speeds.
4Enterprise Performance Gains
Organizations such as Scale AI, Ramp, and AI21 post-trained the model for specialized production deployments. Enterprise adaptations achieved up to 40x lower inference costs, 10x lower latency, and 40% higher accuracy. A specialized medical claims deployment improved task accuracy by 44%.
5Model Distillation Heritage
Nemotron 3.5 Super VL was distilled from the larger Nemotron 3 Ultra using SkewReverse loss to retain high-level coding and reasoning capabilities. Post-training leveraged multi-teacher on-policy distillation, where dedicated vision and text teachers scored model outputs for video reasoning, tracking, and alignment.
6Massive Pretraining Scale
The model underwent training on approximately 26 trillion tokens across initial pretraining and continued pretraining stages. It was then refined using 3.8 million post-training samples alongside reinforcement learning across dozens of simulated enterprise environments to reinforce policy adherence and tool usage.
7Indic Language Datasets
Continued pretraining incorporated 10 billion Indic tokens into the model base. In parallel, NVIDIA released a dataset comprising approximately 2.8 trillion tokens across 22 Indian languages developed in collaboration with IIT Madras and its AI4Bharat lab, alongside open datasets on Hugging Face.
8Advanced Inference Stack
The release supports NVFP4 quantization using 4-bit weights and activations on NVIDIA Blackwell hardware, paired with quantization-aware distillation. Inference optimization features LiLiCorr speculative decoding for up to 2.3x faster token generation on DGX Spark, multi-token prediction, and attention-Mamba context reuse.
9Benchmark Evaluation Results
Super VL recorded 63.44% on Hermes Bench and 87.05% on LangChain behavioral evaluations, surpassing Nemotron 3 Ultra. On the CharXiv chart-reasoning benchmark, it reached 80.8%, rising to 84% with Python code execution, while independent EXO Labs testing placed it on the Pareto frontier.
10Edge-Tier Bolt Deployments
NVIDIA complemented the release with the Nemotron 3.5 Bolt family for local hardware. Bolt 74B-A9B 24GB uses native 2-bit quantization to run locally across RTX workstations, while Bolt E4B 4GB targets lightweight physical and conversational interfaces on Jetson and PC platforms.
11Ecosystem Collaboration Partners
Development was conducted with Nemotron Coalition partners, integrating mid-training datasets from Perplexity and teacher-model environments from Prime Intellect. The models support deployment through NVIDIA NeMo, NIM microservices, and community runtimes including llama.cpp, Ollama, vLLM, Unsloth, and LM Studio.
12Open-Weight Strategic Shift
This launch represents a broader competitive push by Western developers to deliver performant open-weight foundation models against international competitors. By providing customizable weights and open training datasets, NVIDIA enables organizations to retain infrastructure sovereignty while adapting models to proprietary operational workflows.