NVIDIA releases Nemotron 3.5 Super VL: an open 120B model that beats its own flagship on agent tasks

NVIDIA released Nemotron 3.5 Super VL on Thursday — an open 120-billion-parameter multimodal model for specialized AI agents that outscores the company’s own Nemotron 3 Ultra on agent benchmarks and, once post-trained on company data, beats Claude Opus 5 and GPT-6.1 Sol on domain tasks, per NVIDIA’s data.

A smaller heavyweight

Super VL is a 120-billion-parameter mixture-of-experts model that activates 12 billion parameters per token, distilled from NVIDIA's largest reasoning model, Nemotron 3 Ultra — using a technique NVIDIA calls SkewReverse loss during distillation to carry over Ultra's coding ability. The weights are open (base, post-trained, and NVFP4-quantized checkpoints), and NVIDIA says the OpenMDW license permits commercial use, fine-tuning, and redistribution. You can pull the weights from Hugging Face and ModelScope, or try it hosted via OpenRouter or NVIDIA's NIM microservice.

The headline claim: NVIDIA reports Super VL matches Ultra-level accuracy on multimodal agentic tasks at roughly one-fifth the parameter count — and actually beats Ultra on two agent harnesses, scoring 63.44% on Hermes Bench and 87.05% on LangChain behavioral evaluations versus Ultra's 56.33% and 84.3%.

Built to read the paperwork

Super VL folds reasoning, instruction following, and tool use into one model that understands text, images, documents, and video — so an agent no longer has to shuttle context between a language model and a separate vision model. NVIDIA's vision-language training targets what enterprise agents actually meet: OCR, tables, charts, and structured information inside web pages and PDFs.

On CharXiv, a chart-reasoning benchmark, NVIDIA reports 80.8% — versus 63% for its smaller Nemotron 3 Nano Omni — rising to 84% when the model writes Python to analyze the chart. NVIDIA also calls it one of the top open models on OCRBench v2, and says an independent evaluation by EXO Labs placed it on the Pareto frontier for accuracy versus task-completion time on a DGX Spark.

The speed story, with footnotes

NVIDIA describes Super VL as up to five times faster than other open models in its class for agentic tasks, reporting over 600 tokens per second on Artificial Analysis's agentic benchmarks — measured, per VentureBeat, on four Blackwell B200 GPUs. As VentureBeat notes, the chart places Super VL at about 26 on the composite Intelligence Index — below open models like GLM-5.3-Flash and Qwen3.8-Flash-Next — so the 5x is generation speed at that score, not a blanket claim of superiority. Output speed still matters: an agent that makes thousands of calls over a long workflow finishes meaningfully faster.

Under the hood, two efficiency features do the work. LiLiCorr, an open speculative-decoding technique, delivers up to 2.3x faster token generation on DGX Spark by evaluating draft tokens jointly instead of independently. And multi-token prediction gives the model an internal drafter that proposes several tokens per forward pass, while serving optimizations preserve both attention caches and Mamba recurrent states across repeated agent calls — fewer repeated computations when your agent re-reads the same policies and tool definitions every turn.

Post-trained, it beats the flagships — on specific tasks

The most striking numbers are not for the base model. NVIDIA says dozens of organizations are already post-training Super VL on their own data, tools, and policies, and the gains are dramatic: Scale AI post-trained it for Skydance's vendor-credit finance agent, cutting inference cost 40x and latency 5x; Ramp is converting receipt images into structured JSON with up to 50% higher throughput and 10x lower latency; AI21's post-train for Alice improved threat-intelligence accuracy by 40%.

NVIDIA's chart shows customized Super VL variants outscoring rival flagships on five specialized evaluations — including a medical-claims task where accuracy rose from roughly 28% to 72% after post-training versus about 38% for Claude Opus 5, and security attack resilience rising from 35% to 76% versus about 66% for GPT-6.1 Sol. The crucial caveat: these are different customized models on different tasks, NVIDIA has not published task definitions or testing protocols, and VentureBeat rightly flags that buyers should test the approach against their own work rather than take the chart at face value.

Two siblings for the edge

Super VL arrived with two smaller open siblings. Nemotron 3.5 Bolt 74B-A9B uses native 2-bit quantization to fit in 24GB of memory — built for local agents on RTX and DGX Spark hardware, and which NVIDIA calls the top open model in its class on PinchBench. Bolt E4B 4GB, aimed at conversational agents on Jetson and RTX cards, follows on October 23. (We covered the Bolt launch and NVIDIA's PAIR router exit from beta earlier today.)

What it means

The pitch is a shift in how NVIDIA sells open models: not "download the base weights and benchmark them," but "download the base weights, post-train on your documents and procedures, and deploy under your own control." NVIDIA is shipping the raw material alongside it — training recipes, open datasets spanning agentic tool use and finance, 45 vision-language datasets, and roughly 100 reinforcement-learning environments in NeMo Gym. With platform integrations from Atlassian, CrowdStrike, Dassault Systèmes, Perplexity, Salesforce, and Synopsys already announced, the question for enterprise developers is a concrete one: can a post-trained Super VL read your documents and follow your procedures well enough to justify the GPU bill? NVIDIA even quotes the production cost: an NVIDIA AI Enterprise license of $4,500 per GPU per year — software alone, before the hardware.

Sources: NVIDIA developer blog: NVIDIA Nemotron 3.5 Super VL Powers Specialized Agents with Multimodal Reasoning (Oct 15); VentureBeat: NVIDIA's new, open Nemotron 3.5 Super VL with post-training beats Claude Opus 5, GPT-6.1 Sol on specialized task benchmarks (Oct 15); Unite.AI: NVIDIA Releases Open Nemotron 3.5 Super VL and Bolt Edge Models (Oct 15).