NVIDIA Nemotron.
Lightning-fast on PrimaLabs.

Enterprise work arrives in documents, screenshots, charts, and tables. An agent needs to interpret that information and decide what to do next.

Nemotron 3.5 Super VL combines visual understanding with reasoning and tool use. NVIDIA describes it as a 120B-parameter model with 12B active parameters per token. It is built for specialized workflows that need to connect evidence across text and images.

Super VL reaches 589 output tokens per second at 1,000 input tokens, leading the seven-provider chart in this Artificial Analysis snapshot. At 10,000 input tokens, PrimaLabs delivers 567.2 tokens per second. The endpoint supports a full one-million-token context window.

Super VL, measured across providers

Original Artificial Analysis charts

1k input10k input

Artificial Analysis chart

Close

Scroll horizontally to see all provider details.

Artificial Analysis · Captured 22 October 2026. The table reproduces the rounded speed values shown in the original charts and is ordered by 10k output speed. All benchmarked providers are shown. Context is endpoint capacity, not benchmark input length. Provider configurations differ; live results may change.

Super VL’s first answer arrives in 4.35 seconds at 10,000 input tokens in this benchmark snapshot. The full one-million-token context is available for workflows that need substantial source material.

Explore Super VL benchmarks

Nemotron 3.5 Lightning is the smaller, text-only model in the launch. NVIDIA describes it as 30B parameters with 3B active per token, designed for high-volume execution inside agent workflows.

Use it where a smaller model meets your quality requirements: extracting fields, formatting results, generating tool calls, or handling repeated sub-agent steps. Super VL can handle work that depends on visual evidence, while Lightning takes on suitable text tasks.

PrimaLabs leads Artificial Analysis’s six-provider Lightning comparison at both prompt sizes: 533 output tokens per second at 1,000 input tokens and 560.3 at 10,000 input tokens. At 10k input, that is about 22% faster output than the next-fastest provider. We also serve Lightning’s full one-million-token context.

Lightning, measured across six providers

Original Artificial Analysis charts

1k input10k input

Artificial Analysis chart

Close

Scroll horizontally to see all provider details.

Artificial Analysis · Captured 22 October 2026. The table reproduces the rounded speed values shown in the original charts and is ordered by 10k output speed. All benchmarked providers are shown. Context is endpoint capacity, not benchmark input length. Provider configurations differ; live results may change.

PrimaLabs also has the lowest time to first answer token in the 10k Lightning comparison: 4.40 seconds. PrimaLabs and Nebius offer the full one-million-token context; the other four listed providers are capped at 262K.

Faster generation can shorten repeated agent steps. Total workflow time also depends on reasoning length, tool execution, retries, and whether the model gets the task right.

Explore Lightning benchmarks

At GTC Berlin on October 21, NVIDIA CEO Jensen Huang argued that enterprises can build specialized AI by adapting an open model with their own data.

“You could start with an open model, something that’s incredibly good like Nemotron, and with your own data, post-train it for your use case”

That is the opportunity behind this launch. Super VL gives teams a starting point for visual workflows; Lightning handles frequent text steps. PrimaLabs serverless lets developers evaluate both on real workloads, while open weights preserve a path to models tailored to their own documents, terminology, and tools.

• Open weights. Both models use OpenMDW-1.1, which permits commercial use subject to its terms.
• Released training resources. NVIDIA publishes datasets and recipes that teams can inspect and use when adapting the models.
• Room to adapt. Evaluate on a hosted endpoint, then use the results to decide whether fine-tuning or a different deployment fits your requirements.

For us, sovereign AI means control over those decisions. A model’s developer, the environment where it runs, and the way customer data is handled are separate considerations. Open weights give teams more choice over how to put them together.

Less than a week after NVIDIA’s October 15 Super VL release, the model is available on PrimaLabs and independently benchmarked. Our speed of implementation helps teams put new open models to work while the release is still new.

Serverless access lets developers start evaluating without provisioning inference hardware. Measured output speed and full context windows give them concrete serving characteristics to test against their own workloads.

Choose Super VL for workflows that need visual understanding and multistep reasoning. Test Lightning on frequent text tasks. Measure quality, total task time, and cost per successful result, then build around the combination that works.

Sources and methodology: Original Artificial Analysis charts captured 22 October 2026 using single requests at 1,000 and 10,000 input tokens. Chart and table speeds are rounded as displayed by AA; the hero uses the 10k summary values of 567.2 tokens/s for Super VL and 560.3 tokens/s for Lightning. Lightning’s approximately 22% advantage compares 560.3 with 460.0 tokens/s. First-answer latency figures use the 10k workload. Context capacity is separate from measured speed. PrimaLabs serves NVFP4; provider configurations differ. All seven Super VL and six Lightning providers are included. Model specifications and release details are from NVIDIA. Rankings can change. Jensen Huang’s remarks are from NVIDIA’s October 21 GTC Berlin keynote report.

Super VL model card · Lightning model card · NVIDIA Super VL launch