NVIDIA Releases Nemotron 3.5 Super VL Model

NVIDIA has released Nemotron 3.5 Super VL, an open, customizable multimodal model aimed at specialised agent workflows. In a Technical Blog post dated 15 October 2026 by Chris Alexiuk and Chintan Patel, NVIDIA describes a 120-billion-parameter mixture-of-experts (MoE) architecture that activates 12 billion parameters per token, combining reasoning, instruction following and tool use with text, image, document and video understanding in one model.

On a single GPU, NVIDIA says the model delivers up to 5x higher throughput for agentic tasks compared with other open models in its size category. That throughput figure — and every other performance number in this post — is NVIDIA’s claim unless attributed to a named partner.

What the model is for

NVIDIA positions Super VL for always-on agents that juggle longer enterprise workflows: reading a document, checking sources, following policy and calling tools. Example workflows listed include coding and error analysis, web research, document intelligence (tables, charts, layouts), cybersecurity triage and video understanding.

Inside a model stack, NVIDIA says Super VL handles specialised multistep work that needs reasoning plus vision, while smaller Nemotron 3.5 Lightning models cover high-volume text and Nemotron 3 Ultra covers complex planning. Super VL was distilled from Nemotron 3 Ultra.

Benchmarks — NVIDIA’s evaluations

On CharXiv, a chart-reasoning benchmark, NVIDIA says the model scored 80.8%, compared with 63% for Nemotron 3 Nano Omni, and 84% with Python code interpretation. On Hermes Bench it scores 63.44% versus 56.33% for Nemotron 3 Ultra; on LangChain behavioural evals, 87.05% versus 84.3%. NVIDIA also says the model is among the top open entries on OCRBench v2 and outperforms Nemotron 3 Super on SWE-bench and Terminal-Bench across multiple setups.

These are NVIDIA-run or NVIDIA-cited evaluations, not independent third-party audits published with the blog.

Partner post-training claims

Organisations named as post-training Super VL for production workflows include:

Scale AI for automated vendor credit processing at Skydance — Scale says inference cost for this workload “will decrease by 40x and latency by 5x” (projected)
Ramp for converting receipt images to structured JSON — Ramp “expects up to 50% higher throughput and 10x lower latency”
AI21 for Alice’s threat intelligence and vulnerability analysis — “improving accuracy by 40%“

Present those as partner claims, not verified TSN measurements. Ramp’s role here is product customisation of an open model; TSN separately covered Ramp’s US$1.85 billion raise at a US$60 billion valuation.

Open release package

NVIDIA says the release includes base, post-trained and NVFP4-quantized checkpoints with training recipes, data and reinforcement-learning environments. Open datasets on Hugging Face include agentic tool-use data, finance reasoning sets, vision-language datasets and about 2.8 trillion tokens across 22 Indian languages, developed with IIT Madras and AI4Bharat. Hosted access is via OpenRouter; NVIDIA NIM packages production deployment; the model is optimised for DGX Spark.

Serving claims include LiLiCorr speculative decoding at up to 2.3x faster token generation on DGX Spark, and NVIDIA Dynamo on 16 B200 GPUs serving up to 8x as many users while cutting first-token latency by 85% at a stated 50 tokens-per-second-per-user interactivity — again NVIDIA’s figures.

What this post does not claim

Some secondary coverage has framed specialised post-trained results as beating named frontier chat models such as Claude Opus 5 or GPT-6.1 Sol. Do not treat that as a general claim about the base Super VL release. NVIDIA’s blog focuses on open-model comparisons, its own tables and partner workflow claims. The blog also does not state a licence in the materials checked for this draft, so none is reported here — check Hugging Face model cards before redistributing weights.

Why it matters

Open multimodal MoEs that keep active parameters low are competing for the “local or single-GPU agent” slot that NVIDIA has been seeding with RTX Spark / DGX Spark tooling and Run:ai open-sourcing. Super VL extends that stack into document-and-video agents with a post-training path through NeMo.