Nemotron 3.5 Super VL: A 26 Score At 25 Cents Per Task

🔍 Read the full analysis: Nemotron 3.5 Super VL: A 26 Score At 25 Cents Per Task on ThorstenMeyerAI.com

Before you orderOffer from Amazon

Get hardware and tech essentials 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, with Artificial Analysis reporting an Intelligence Index score of 26.5 and an estimated cost of $0.25 per index task. The 120-billion-parameter open-weight model is designed for text and image input, and its reported speed and commercial-use licensing may appeal to teams weighing deployment control against raw benchmark performance.

The model uses a mixture-of-experts design with 120 billion total parameters and 12 billion active parameters. NVIDIA released its weights in NVFP4, a four-bit format, and the stated context window is one million tokens. Its OpenMDW-1.1 license allows commercial use, according to the source material; the terms and any deployment-specific conditions should be checked directly before use.

Artificial Analysis lists API rates of $0.20 per million input tokens and $0.575 per million output tokens, with cached input priced at $0.055 per million tokens. Its estimated $0.25 cost per Intelligence Index task is substantially below the $1.64 listed for Nemotron 3 Super, the predecessor, while the reported index score rises from 13 to 26. Those task-cost figures reflect the evaluation setup, not a guaranteed bill for other workloads.

Speed varies by measurement. A pre-release DeepInfra endpoint running the NVFP4 weights produced about 630 output tokens per second on prompts of 100,000 tokens, according to Artificial Analysis. Across six providers, the reported median was 307 tokens per second on 100,000-token prompts, with results ranging from 106 to 534. Provider, hardware and workload can all affect actual performance.

Speed, Cost and Model Trade-offs

The release puts a relatively inexpensive, commercially usable model into a part of the market where performance is only one deployment concern. Teams may value running weights on their own NVIDIA hardware, controlling where data is processed, or reducing reliance on a particular hosted provider. Those benefits can matter even when another model earns a higher benchmark score.

However, the reported index score does not make Super VL a leading option for raw intelligence per dollar. The source comparison lists MiMo-V2.6-Pro at 46 for $0.13 per task, MiMo-V2.6-Flash at 38 for $0.06, and Inkling Small at 26 for $0.09. These comparisons use Artificial Analysis figures and should be read alongside differences in licensing, availability, task mix and deployment requirements. A benchmark task is not a complete measure of production value.

For operators, the pricing profile also makes input usage important. The source reports that about 89% of the $0.25 task estimate is input cost, including 69% from cache writes, while output costs less than three cents. An agent that repeatedly processes large documents or repositories may therefore have costs that depend more on its input and caching patterns than on the output-token rate alone.

How the Benchmark Figures Compare

Artificial Analysis places Super VL at the top of its 40-billion-to-150-billion-parameter open-weight size class, which contains 67 models. That ranking is narrower than a claim that it leads all open-weight models: the source estimates it ranks about 18th when all open-weight models are counted once. The size class and comparison set matter when interpreting the result.

The reported speed advantage also depends on the test. The 630-token-per-second figure came from one pre-release DeepInfra endpoint; the six-provider median for 100,000-token prompts was about half that amount. Artificial Analysis reports a median time to first token of 0.89 seconds, but says the first answer token takes about 11 seconds when the model’s reasoning period is included. It estimates about 13.6 seconds for a 500-token answer. These figures describe different timing measures and should not be treated as interchangeable.

The model’s evaluation results are uneven. The source reports a GDPval-AA score of 1,162 Elo, up from 490 for Nemotron 3 Super, and an AA-Briefcase score of 999 versus 876 for Nemotron 3 Ultra. AutomationBench-AA is reported at 14.8%, compared with 3.0% for Ultra. The AA-Omniscience hallucination rate is reported to fall from 87% to 36%. These are benchmark results, not assurances about accuracy in a particular application.

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

— Thorsten Meyer, ThorstenMeyerAI.com

Production Results Still Vary

The published figures do not establish how Nemotron 3.5 Super VL will perform across different production workloads, hardware configurations or providers. The 630-token-per-second result came from a pre-release endpoint, and the supplied material does not identify the exact conditions behind every provider result. Teams would need to test their intended endpoint, prompt lengths and response requirements.

The $0.25 task estimate is tied to Artificial Analysis’s evaluation workload; it is not a universal cost per user request. Actual spending will depend on input and output volume, cache behavior, provider prices and how often an application repeats or expands tasks. The supplied material also does not provide full benchmark methodology, confidence intervals or independent validation of the reported figures.

It remains unclear how well the benchmark gains translate to a given organization’s software, and whether local deployment will deliver the expected savings after hardware and operating costs. Comparisons with other models may also change as providers revise prices, models or evaluation results.

Test the Intended Deployment

The next practical step for teams considering the model is to benchmark their own workloads against the provider or self-hosted setup they intend to use. Tests should track task accuracy, latency, throughput and total input and output costs, including cached inputs. For image-heavy applications, teams should also evaluate the relevant visual tasks rather than relying only on an aggregate index score.

Organizations should review the OpenMDW-1.1 license and deployment terms before commercial use, and compare the operational requirements of self-hosting with the available API options. The source material does not give a date for further evaluation updates or establish whether current provider performance will remain stable. Pricing, model availability and benchmark rankings may change.

Key Questions

What is Nemotron 3.5 Super VL?

It is an NVIDIA open-weight model released October 15, 2026. It accepts text and images, produces text, and has 120 billion total parameters, with 12 billion active in its mixture-of-experts design.

What does the score of 26 mean?

Artificial Analysis reports an Intelligence Index score of 26.5. The source says it ranks first among 67 open-weight models in the 40-billion-to-150-billion-parameter class, but about 18th among open-weight models overall. The score is a benchmark, not a direct prediction of performance on every task.

Does each task cost $0.25?

No. $0.25 is an estimated cost per Artificial Analysis index task under its evaluation setup. Costs for other applications will vary with token use, caching, provider rates and workload design.

Is it the fastest model available?

The source reports about 630 output tokens per second on 100,000-token prompts from one pre-release endpoint. Across six providers, the median for that prompt length was 307 tokens per second. The supplied evidence does not support a general claim that it is the fastest model across all providers and conditions.

Can businesses use the model commercially?

The source says its OpenMDW-1.1 license allows commercial use. Businesses should review the license text and their deployment arrangement to confirm the terms apply to their intended use.

Source: ThorstenMeyerAI.com