NVIDIA introduced Nemotron 3.5 Super VL as an open, customizable multimodal model designed for always‑on enterprise agents that must execute longer, multi‑step workflows. Modern single requests frequently require reading documents, checking multiple sources, following company policies and calling external tools; Nemotron 3.5 Super VL is built to handle these combined requirements.
Architecture and performance
Nemotron 3.5 Super VL is a 120 billion‑parameter mixture‑of‑experts (MoE) model that activates about 12 billion parameters per token to balance capacity and inference efficiency. NVIDIA reports that on a single GPU the model can deliver up to 5× higher throughput on agentic tasks compared with other open models in the same size category. On a Baseten endpoint benchmark for agentic tasks, the model achieved over 600 TPS (transactions per second).
The model unifies reasoning, instruction following and tool use with text, image, document and video understanding, enabling teams to address combined text and visual tasks without deploying separate text and vision‑language models or passing context between them.
Typical agent workflows
Nemotron 3.5 Super VL is intended for workflows that require both task understanding and knowledge of an operating environment. Example use cases include:
- Coding: test and validate code, analyze errors and recommend fixes.
- Web research: search websites, connect information and verify answers.
- Document intelligence: extract data from tables, charts and page layouts to create summaries and reports.
- Cybersecurity: identify vulnerabilities and triage threats.
- Video understanding: extract information for search, content review and incident analysis.
In multi‑model systems, Nemotron 3.5 Super VL handles specialized, multi‑step jobs needing combined reasoning, tool use and multimodal understanding, while smaller models such as Nemotron 3.5 Lightning serve high‑volume text tasks and frontier models like Nemotron 3 Ultra address complex planning and orchestration.
Customization and partner examples
Organizations can post‑train Nemotron 3.5 Super VL with NVIDIA NeMo, using their own data, tools, policies and workflows. NVIDIA provides NeMo AutoModel for supervised fine‑tuning (SFT) and low‑rank adaptation (LoRA); reinforcement learning (RL) can improve performance in agentic environments.
Examples from partners who poszttanított the model include:
- Scale AI: post‑trained Nemotron 3.5 Super VL for automated vendor credit processing at Skydance Corporation; inference cost is expected to drop 40× and latency to improve 5× for that workload.
- Ramp: converting receipt images to structured JSON with merchant, date and payment fields; Ramp expects up to 50% higher throughput and 10× lower latency.
- AI21: post‑training for Alice’s threat intelligence and vulnerability analysis improved accuracy by 40%.
NVIDIA also releases open datasets on Hugging Face to supplement proprietary data for SFT and RL, including agentic tool‑use and conversational data, finance reasoning/tool‑calling datasets, instruction and chat data, long‑context transduction and translation data, fairness datasets and 45 vision‑language datasets. For multimodal data, NeMo Data Designer provides chart question‑answering and long‑document understanding resources.
The model’s open weights can be adapted and deployed locally, in data centers, or in cloud environments; a usage cookbook on GitHub documents customization procedures.
Multimodal training and benchmarks
The visual training covers charts, tables, OCR, multipage documents and video understanding. A two‑stage SFT pipeline trains language and vision components together, combining visual evidence with reasoning, chat, safety, long‑context data and agent trajectories.
On the CharXiv chart‑reasoning benchmark, Nemotron 3.5 Super VL scored 80.8%, compared with 63% for Nemotron 3 Nano Omni; with Python code interpretation it reached 84%. The model is also among the top open models on the OCRBench v2 leaderboard.
Accuracy and efficiency for agent tasks
Nemotron 3.5 Super VL has been evaluated across Hermes, LangChain, OpenClaw, OpenCode, Pi, Codex and Claude, as well as in providers’ internal agent harnesses. On Hermes Bench and LangChain behavioral evaluations, the model scored higher than Nemotron 3 Ultra:
- Nemotron 3 Ultra: Hermes Bench 56.33%, LangChain Evals 84.3%.
- Nemotron 3.5 Super VL: Hermes Bench 63.44%, LangChain Evals 87.05%.
For always‑on agents, efficiency—throughput, token use, tool calls and steps to reach correctness—determines how much work a deployment can complete. NVIDIA reports that Nemotron 3.5 Super VL’s throughput and token efficiency can deliver up to 4× faster task completion in some setups. An independent EXO Labs evaluation placed the model on the Pareto frontier for accuracy and task completion time on NVIDIA DGX Spark.
Inference optimizations and training techniques
Nemotron 3.5 Super VL was distilled from Nemotron 3 Ultra; Ultra’s coding capabilities and a SkewReverse loss during distillation improved coding accuracy. Post‑training used multi‑teacher on‑policy distillation (MOPD) to combine guidance from specialized teacher models across text and vision, and additional training on agent actions, code repositories and software engineering tasks improved agent performance.
Continued pretraining with high‑quality natural and synthetic data added 10 billion Indic tokens from a 12‑language subset. NVIDIA is releasing a full Indic dataset of about 2.8 trillion tokens across 22 Indian languages, developed with IIT Madras and AI4Bharat.
Key inference techniques include:
- LiLiCorr: an open speculative decoding method that evaluates token candidates jointly across a draft to capture dependencies; on DGX Spark it can deliver up to 2.3× faster token generation than decoding without speculation.
- NVFP4 and quantization‑aware distillation (QAD): a 4‑bit weight/activation recipe for NVIDIA Blackwell architecture that reduces model size while QAD helps retain accuracy; QAD can also store KV cache in NVFP4 without sacrificing targets.
- Multi‑token prediction (MTP): an internal drafter that proposes multiple tokens in one forward pass, reducing verification steps. A dedicated MTP training phase aligns the drafter with the final model.
- Context reuse and prefix caching: reuse computation for repeated system instructions, tool definitions and conversation history; serving optimizations preserve attention caches and Mamba recurrent states to minimize replay.
Deployment, scaling and tools
NVIDIA NIM packages models with optimized inference engines, validated configurations and standard APIs for cloud and data center deployment; enterprise support is available through NVIDIA AI Enterprise.
NVIDIA Dynamo enables distributed inference optimizations such as KV‑cache‑aware routing and KV cache management. In one test on 16 NVIDIA B200 GPUs at 50 tokens/second per user, KV‑aware routing with session affinity and KV cache offload supported up to 8× as many users while reducing first‑token latency by 85%.
Availability and ecosystem
The release includes base, post‑trained and NVFP4‑quantized checkpoints with training recipes, data and RL environments. Detailed technical information is in the Nemotron 3.5 technical report; deployment instructions are available via Hugging Face, ModelScope, and hosted access through OpenRouter.
A broad partner ecosystem supports post‑training, inference software, agent frameworks, cloud platforms, system integrators, hardware and hosted inference providers to help organizations customize and deploy Nemotron 3.5 Super VL for domain‑specific workflows.
Summary
Nemotron 3.5 Super VL is an open, multimodal MoE model aimed at enterprise, always‑on agent workloads, combining multimodal understanding, reasoning and tool use. NVIDIA provides tooling, datasets and deployment options (including quantized checkpoints and distributed inference recipes) so organizations can adapt the model to their data, policies and infrastructure.
