Nemotron 3.5 Super VL’s 26 Score, Explained

🔍 Read the full analysis: Nemotron 3.5 Super VL’s 26 Score, Explained on ThorstenMeyerAI.com

Buying for a business?Offer from Amazon

Get business pricing on tech for your team

• Business-only prices and quantity discounts
• Tax-exempt purchasing
• Multiple users, one account, clear invoices

Create a free Business account →Free to register

As an affiliate, we earn on qualifying purchases.Hide

NVIDIA released Nemotron 3.5 Super VL on October 15, 2026, and Artificial Analysis scored it 26 on its Intelligence Index, twice the 13 recorded for its predecessor, Nemotron 3 Super. The result puts it at the top of the source’s 40B-to-150B open-weight size class, but comparisons in that same source show several models with higher scores at similar or lower estimated cost per task.

The model has 120 billion total parameters, with 12 billion active in its mixture-of-experts design, and accepts text and images while producing text. NVIDIA released weights in NVFP4, a four-bit format. Its stated context window is one million tokens. The source lists the OpenMDW-1.1 license as allowing commercial use, six providers, and API prices of $0.20 per million input tokens and $0.575 per million output tokens. Cached input is listed at $0.055 per million tokens.

Artificial Analysis estimates $0.25 per Intelligence Index task, compared with $1.64 for Nemotron 3 Super. Its launch write-up attributes part of the reduction to lower output volume: about 46,000 output tokens per index task, compared with 114,000 for the predecessor. The source’s cost breakdown says input accounted for about 89% of the task bill, including 69% from cache writes. Those figures describe the benchmark’s task profile; an individual deployment’s costs will depend on its prompts, caching and usage.

Speed varies by measurement. A pre-release DeepInfra endpoint serving NVFP4 weights produced about 630 output tokens per second with 100,000-token prompts, which Artificial Analysis described as its fastest measurement at that length among models scoring at least 20. Across six providers, however, the source reports median output speeds of 196 tokens per second on standard prompts and 307 with 100,000-token prompts. The reported provider range is 106 to 534 tokens per second at the 5th and 95th percentiles.

Where Speed Beats Score

The score of 26 is a useful comparison point, but it does not by itself settle which model suits a team. In the supplied source’s task-cost comparisons, Xiaomi’s MiMo-V2.6-Pro scores 46 at $0.13 per task, while MiMo-V2.6-Flash scores 38 at $0.06. DeepSeek V4.1 Flash is listed at 39 for $0.27, and GLM-5.3-Flash at 42 for $0.28. Inkling Small matches Super VL’s 26 score at a listed $0.09 per task. These are source-reported benchmark and cost estimates; they do not establish performance on every real workload.

Super VL’s case instead combines throughput, latency, commercial-use terms and the option to run the weights on NVIDIA hardware. Those features can matter to teams with strict deployment or licensing requirements. The source reports a median time to first token of 0.89 seconds, compared with 1.76 seconds for Inkling Small and 1.23 seconds for Nemotron 3 Ultra. It also says the first answer token takes about 11 seconds because the model reasons before responding, and a 500-token answer takes about 13.6 seconds end to end. That distinction matters for interactive use: low initial latency does not mean the full answer arrives immediately.

The source also identifies knowledge work and hallucination reduction as areas of improvement over the predecessor. Its GDPval-AA score rose to 1,162 Elo from 490, and its AA-Omniscience hallucination rate fell from 87% to 36%. These are benchmark results reported in the source, not guarantees about accuracy or error rates in a particular application.

How the Benchmark Ranks

Artificial Analysis’ headline ranking compares open-weight models in the same total-parameter range. The source reports Super VL as first among 67 models in its 40B-to-150B class; across all open-weight models, it places the model around 18th when each model is counted once. The distinction explains how a model can lead its size class without leading the wider field.

The source describes Super VL as the top-scoring permissively licensed US open-weight model. That wording depends on both the geographic category and the license criteria. Meta’s Muse Spark Open, released on the same day according to the source, scores 39 but uses a license the source says allows commercial use with restrictions. Teams should review the relevant license terms for their intended deployment rather than treating a category label as legal advice.

The release follows NVIDIA’s Nemotron 3 Super, whose reported Intelligence Index score was 13. Alongside the higher score, the source reports improvements in GDPval-AA and AutomationBench-AA: the latter reached 14.8%, compared with 3.0% for Nemotron 3 Ultra. The source characterizes that as roughly five times higher, while also noting the absolute result remains low. Benchmark gains provide context for the release, but do not show how the model will perform across a team’s own tasks.

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

— Thorsten Meyer, ThorstenMeyerAI.com

What the Score Cannot Show

The supplied source does not establish how Super VL performs on a particular company’s workload, or whether the reported provider prices and speed measurements will hold for every deployment. The 630-token-per-second result came from one pre-release endpoint; the cross-provider median and range show that endpoint choice changes observed throughput. Teams would need to measure the provider and configuration they plan to use.

The source also does not specify the methodology behind every cost estimate in enough detail to calculate an individual user’s bill from the headline task price. Its cost breakdown describes one benchmark profile, where input and cache writes made up most of the estimate. Actual prompt lengths, cache behavior, output size and provider pricing can change the result. The reported benchmarks likewise do not establish accuracy, reliability or hallucination rates for every use case.

Finally, the article’s characterization of license permissiveness and US origin depends on criteria not fully detailed in the supplied material. Commercial use is stated as allowed under OpenMDW-1.1, but organizations still need to check the license text against their own requirements. The source does not provide a full account of restrictions or deployment conditions.

Test the Intended Deployment

The next practical step for prospective users is to benchmark the model on their own tasks and check the actual provider, quantization and hardware configuration. That assessment should include task quality, end-to-end response time, token use and cached-input behavior, since the source’s figures show each can affect cost or speed.

Teams weighing Super VL against alternatives can compare models on the same prompts and workflows, then review the applicable license terms before deployment. The supplied source does not identify a scheduled follow-up release or a next benchmark milestone. Further provider measurements and independent evaluations could clarify whether the reported speed and task-cost advantages hold across more deployments.

Key Questions

What does Nemotron 3.5 Super VL’s score of 26 mean?

It is the model’s reported Artificial Analysis Intelligence Index score. The supplied source says it is twice Nemotron 3 Super’s score of 13 and ranks first in the 40B-to-150B open-weight size class. It does not mean the model will score 26 on every task or perform best in every deployment.

Is Super VL the best-value open model?

The comparisons in the source do not support that conclusion for raw Intelligence Index score per task cost. Several listed open models score higher at similar or lower estimated task costs, and Inkling Small matches 26 for a lower listed cost. Super VL may still suit teams that prioritize its speed, license terms or self-hosting option.

How fast is the model?

A pre-release DeepInfra endpoint reached about 630 output tokens per second on 100,000-token prompts. Across six providers, the reported median was 307 tokens per second for those prompts, with a 106-to-534 range. Results depend on the provider and setup.

What is the estimated cost per task?

Artificial Analysis estimates $0.25 per Intelligence Index task. The source says about 89% of that estimate came from input, including cache writes. A real workload may cost more or less depending on its prompts, caching, output and provider.

What remains to be checked before using it?

Prospective users should test the model on their own tasks, measure their chosen endpoint’s speed and costs, and review the OpenMDW-1.1 terms for their deployment. The source’s benchmarks and price estimates do not guarantee results for a particular application.

Source: ThorstenMeyerAI.com