Nemotron 3.5 Super VL: What 25 Cents A Task Gets You

🔍 Read the full analysis: Nemotron 3.5 Super VL: What 25 Cents A Task Gets You on ThorstenMeyerAI.com

Buying for a business?Offer from Amazon

Get business pricing on monitors, keyboards and dev gear

• 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 rated it 26.5 on its Intelligence Index, with an estimated cost of $0.25 per index task. The 120-billion-parameter, image-and-text model offers commercial use under its stated license and fast inference on some endpoints, but its benchmark score trails several competing models available at similar or lower task costs.

Artificial Analysis places Super VL first among 67 open-weight models in its 40-billion-to-150-billion-parameter size class. Across all open-weight models counted once, the source material puts it at about 18th. NVIDIA’s model has 120 billion total parameters and 12 billion active parameters, uses a mixture-of-experts design, accepts images and text, and has a stated context window of one million tokens. Its released weights use NVFP4, a four-bit format.

Artificial Analysis lists API prices of $0.20 per million input tokens and $0.575 per million output tokens, with cached input priced at $0.055 per million. The source estimates $0.25 per Intelligence Index task, compared with $1.64 for Nemotron 3 Super. It attributes the lower task cost partly to fewer output tokens per evaluation task and cheap cached input; its cost breakdown says roughly 89% of the bill comes from input, including cache writes.

Speed depends on where the model runs. A pre-release DeepInfra endpoint recorded about 630 output tokens per second with 100,000-token prompts, which Artificial Analysis described as its fastest measurement at that prompt length among models scoring at least 20. Across six providers, however, the reported median was 196 tokens per second on standard prompts and 307 with 100,000-token prompts. The first token arrived in a median of 0.89 seconds, but the model’s reasoning can make the first answer token take about 11 seconds.

Where the $0.25 Task Fits

The figures point to a model that may suit teams prioritizing throughput, local deployment and commercial licensing over the highest benchmark score per dollar. The source compares Super VL with MiMo-V2.6-Pro, which scores 46 for $0.13 per task, and Inkling Small, which scores 26 for $0.09. It also lists DeepSeek V4.1 Flash at 39 for $0.27 and GLM-5.3-Flash at 42 for $0.28. These are Artificial Analysis comparisons, not guarantees of performance on a particular workload.

Those comparisons make the distinction between a cheap token price and good value per task important. The $0.25 estimate is based on the benchmark’s task mix; actual bills will depend on prompt length, caching, output volume and the provider. For applications that repeatedly supply large documents or code repositories, input and cache costs may matter more than output pricing.

The model’s headline speed also needs a practical test. The 630-token-per-second result came from one pre-release endpoint, while provider-wide figures range from 106 tokens per second at the fifth percentile to 534 at the 95th percentile. The reported 11-second wait for a first answer token could work for background agents and batch processes, but may feel slow in interactive chat. Teams should test the specific endpoint and workflow they plan to use.

Benchmarks Behind the Launch

Artificial Analysis published its figures on the same day NVIDIA released the model. Its Intelligence Index score doubled from 13 for Nemotron 3 Super to 26.5 for Super VL, while the estimated cost per index task fell from $1.64 to $0.25. The source also reports substantial differences in selected task results: GDPval-AA, a measure of economically useful knowledge work, rose from 490 Elo for the predecessor to 1,162 for Super VL.

Other reported results include 14.8% on AutomationBench-AA, compared with 3.0% for Nemotron 3 Ultra, and an AA-Omniscience hallucination rate of 36%, down from 87% for Nemotron 3 Super. The source describes the automation result as an improvement but notes the absolute score remains low. These benchmark results measure defined test sets and should not be treated as proof that the model will perform similarly across every business task.

The licensing distinction also shapes the comparison. The source says OpenMDW-1.1 allows commercial use and calls Super VL the leading permissively licensed US open-weight model in its comparison. It contrasts that with Meta’s Muse Spark Open, which it says scored 39 but uses a license with commercial-use restrictions. “Open-weight” does not by itself mean every model has the same license terms, so teams need to review the applicable license for their intended use.

Endpoint and Workload Differences

The published figures do not establish how Super VL will perform on a given company’s prompts, hardware or production load. Provider speed varies substantially, and the 630-token-per-second figure reflects one pre-release endpoint rather than a provider-wide result. It is also not clear from the supplied material how the task-cost estimate would change for workloads with different prompt lengths, cache-hit rates or output requirements.

The benchmark results are attributed to Artificial Analysis, but the source material does not provide the full evaluation methodology or independent replication details. The stated million-token context window is a model specification; it does not show how accuracy or latency changes as prompts approach that limit. The license is identified as OpenMDW-1.1, but organizations should check its full terms rather than infer permissions from the label alone.

Production Tests Will Set the Value

The next practical step for prospective users is to test the provider, hardware and task mix they would actually deploy. That means comparing latency, accuracy, token use and total cost on representative prompts, rather than relying only on the model’s index score or best reported speed. Teams considering self-hosting will also need to assess whether their NVIDIA hardware can serve the NVFP4 weights at the required throughput.

No further release milestone or additional benchmark schedule is specified in the source material. As more providers and production deployments report results, the key question will be whether Super VL’s speed and licensing advantages offset its lower score per dollar for particular workloads.

Key Questions

What is Nemotron 3.5 Super VL?

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

What does “25 cents a task” mean?

The source reports an estimated $0.25 per Artificial Analysis Intelligence Index task. It is a benchmark-based cost estimate, not a fixed price for every real-world request; actual costs depend on token use, caching and provider pricing.

Is it the fastest model available?

The supplied material reports about 630 output tokens per second on one pre-release endpoint with 100,000-token prompts. Across six providers, the median was 307 tokens per second for that prompt length, so the peak result should not be assumed for every deployment.

Does its license allow commercial use?

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

Is it better value than competing models?

Not on the supplied Intelligence Index cost comparisons alone. The source lists several models with higher scores at similar or lower estimated task costs. Super VL’s potential advantages are speed, licensing and self-hosting, which teams should weigh against results on their own workloads.

Source: ThorstenMeyerAI.com