What Do Nemotron 3.5 Super VL’s Three Headline Metrics Mean?

🔍 Read the full analysis: What Do Nemotron 3.5 Super VL’s Three Headline Metrics Mean? on ThorstenMeyerAI.com

Before you orderOffer from Amazon

Get the latest gadgets delivered free with Prime

• Fast, free delivery on millions of items
• Prime Video, Amazon Music and more included
• Member-only deals all year

Start your free Prime trial →Free trial for eligible customers · Cancel anytime

As an affiliate, we earn on qualifying purchases.Hide

NVIDIA released Nemotron 3.5 Super VL on October 15, 2026, and Artificial Analysis reported a score of 26 on its Intelligence Index the same day. The result is double the 13 scored by its predecessor, Nemotron 3 Super, and leads the 40 billion to 150 billion parameter class of open-weight models tracked by the service. The figures describe a model designed to combine speed, relatively low task cost and commercial-use licensing, while leaving a substantial gap between its benchmark score and several similarly priced alternatives.

The model has 120 billion total parameters, with 12 billion active at a time through a mixture-of-experts design, and a one-million-token context window. It accepts text and images and produces text. NVIDIA released weights in NVFP4, a four-bit format. Its listed license, OpenMDW-1.1, allows commercial use, according to the source material. The model can be accessed through six providers or run on NVIDIA hardware.

Artificial Analysis lists API prices of $0.20 per million input tokens and $0.575 per million output tokens, with cached input at $0.055 per million tokens. It estimates an Intelligence Index task costs about $0.25. For comparison, Nemotron 3 Super cost $1.64 per task in the cited figures. The source attributes the lower bill in part to fewer output tokens per index task, about 46,000 versus 114,000 for the predecessor, and cheaper cached input.

The speed figures depend on where the model runs. A pre-release DeepInfra endpoint serving NVFP4 weights produced about 630 output tokens per second with 100,000-token prompts. Across six providers, Artificial Analysis lists a median of 196 tokens per second on its standard prompt and 307 with 100,000-token prompts; results ranged from 106 to 534 tokens per second. Median time to first token was 0.89 seconds, but the source says reasoning before the answer can take about 11 seconds, with a 500-token answer taking roughly 13.6 seconds end to end.

Speed Comes With a Provider Caveat

The difference between the single-endpoint peak and the six-provider median matters for teams budgeting workloads or setting response-time expectations. A result of 630 tokens per second demonstrates what one setup achieved; it does not predict the performance every customer will see. Provider choice changes the measured speed substantially, so the relevant figure is the one from the endpoint a team expects to use.

The task-cost estimate also has a practical implication: input accounts for most of the bill. The source’s breakdown puts about 89% of the $0.25 task cost on input, including 69% from cache writes, while output costs less than three cents. For agents that repeatedly process documents or repositories, prompt and cache behavior may matter more to total cost than the output-token price alone.

A score of 26 is a useful capability signal, but it does not establish best value. The source compares it with MiMo-V2.6-Pro, which scores 46 at $0.13 per task, and Inkling Small, which also scores 26 at $0.09. Super VL’s case instead includes its reported speed, commercial-use license, NVIDIA provenance and self-hosting option. Which of those matter most depends on a team’s operating and legal requirements.

Products Worth Considering

Kindle Paperwhite Signature Edition 32GB (2024) – Auto-adjusting front light, wireless charging, weeks of battery life – Metallic Black

AmazonView Latest Price

LG UltraGear 32GX850A-B 32-inch 4K UHD Glossy OLED Gaming Monitor

AmazonView Latest Price

Bose QuietComfort Wireless Noise-Cancelling Headphones, Cypress Green

AmazonView Latest Price

As an affiliate, we earn on qualifying purchases.

How the Score Compares

The Intelligence Index combines results across evaluations into a single score. In the source’s account, Super VL’s 26.5 figure is rounded to 26 in the headline comparison. Its “number one” position applies to 67 open-weight models in its size class, defined here as 40 billion to 150 billion total parameters. The source places it around 18th among open-weight models overall, counting each model once. Those rankings answer different comparison questions.

The per-evaluation results vary. Artificial Analysis’s GDPval-AA score, which covers economically useful tasks, rose to 1,162 Elo from 490 for Nemotron 3 Super. The source says this exceeds Nemotron 3 Ultra’s 1,013 and Inkling Small’s 1,124. On AA-Briefcase, Super VL scored 999 Elo against 876 for Ultra. AutomationBench-AA was 14.8%, compared with 3.0% for Ultra; the source describes that as about five times higher while also noting the absolute result remains low.

The source also reports an AA-Omniscience hallucination rate of 36%, down from 87% for Nemotron 3 Super. This is a benchmark result, not a guarantee that the model will avoid errors in a particular deployment. Separately, the source says Super VL leads among permissively licensed US open-weight models. Meta’s Muse Spark Open scored 39 in the comparison but uses a license that the source describes as allowing commercial use with restrictions.

““Super VL is a very good engine and an average brain.””

— Thorsten Meyer, ThorstenMeyerAI.com

Deployment Results Still Vary

The reported 630 tokens-per-second result came from one pre-release endpoint, and the available material does not establish that customers will reproduce it on other providers or hardware. Provider-level speed, latency and pricing can differ; the six-provider range already shows a wide spread. The precise performance of a self-hosted deployment will depend on its hardware and configuration, which the cited figures do not specify.

The index is an aggregate and does not establish how the model will perform on a particular company’s tasks, data or safeguards. The supplied account gives selected benchmark results, but does not provide full evaluation methods or enough detail to independently assess each score. The 36% hallucination rate is tied to AA-Omniscience and should not be read as a general error rate for all use cases.

It is also not clear from the supplied figures how costs will compare across real workloads with different prompt lengths, caching patterns, reasoning settings and provider rates. The $0.25 figure refers to an Intelligence Index task under Artificial Analysis’s accounting. Teams should not assume it represents a typical production request.

Test the Intended Workload

Teams considering Super VL can compare their own tasks against the benchmarks and run a limited evaluation on the provider or NVIDIA hardware they expect to use. That should capture full response time, output quality, failure rates and token usage, including cache writes and long input prompts. The published numbers give a starting point for that evaluation, not a replacement for it.

Before deployment, teams will also need to check the terms of OpenMDW-1.1 against their intended use and confirm the hardware and operating costs of self-hosting. Future provider measurements and independent evaluations could clarify how consistently the reported speed and task costs hold across setups. No such follow-up results are included in the material cited here.

Key Questions

What does a score of 26 mean for Nemotron 3.5 Super VL?

It is the model’s reported score on Artificial Analysis’s Intelligence Index. The source says the score is double Nemotron 3 Super’s 13 and ranks first among 67 open-weight models in the 40 billion to 150 billion parameter class. It places Super VL around 18th among open-weight models overall.

Is 630 tokens per second the speed most users should expect?

No single speed is established for all deployments. The 630-token-per-second result came from one pre-release DeepInfra endpoint with 100,000-token prompts. Across six providers, the reported median was 307 tokens per second with prompts of that length, with results ranging from 106 to 534.

How much does an Intelligence Index task cost?

Artificial Analysis estimates about $0.25 per task for Super VL. The source says about 89% of that amount comes from input, including cache writes. Actual costs depend on workload and provider.

Does the license allow commercial use?

The source identifies the model’s license as OpenMDW-1.1 and says commercial use is allowed. Organizations should review the license terms against their intended deployment.

Is Super VL the best-value model in its class?

The cited comparison does not support calling it the best value for raw Intelligence Index score per task. The source lists models with higher scores at similar or lower estimated task costs. Super VL’s potential advantages include speed, its stated license and the option to run it on NVIDIA hardware.

Source: ThorstenMeyerAI.com