What Nemotron 3.5 Super VL Offers For 25 Cents A Task

🔍 Read the full analysis: What Nemotron 3.5 Super VL Offers For 25 Cents A Task 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 reported an Intelligence Index score of 26.5 and an estimated $0.25 per index task. The model pairs a 120-billion-parameter mixture-of-experts design with a commercial-use license and fast inference, but the source report says its benchmark score trails several models available for similar or lower task costs.

Artificial Analysis lists Super VL at 26.5 on its Intelligence Index, roughly twice the score of its predecessor, Nemotron 3 Super, at 13. The model has 120 billion total parameters and 12 billion active, accepts text and images, and has a stated context window of one million tokens. It is released in NVFP4 format, a four-bit representation, and under the OpenMDW-1.1 license, which the report says allows commercial use.

The reported API rates are $0.20 per million input tokens and $0.575 per million output tokens, with cached input listed at $0.055 per million tokens. Artificial Analysis estimates a complete Intelligence Index task costs $0.25, compared with $1.64 for Nemotron 3 Super. The report attributes the lower task cost partly to lower output volume: about 46,000 output tokens per index task, against 114,000 for the predecessor. It says most of the estimated bill is input-related, so actual costs will depend on how much material a workload sends and reuses.

Speed figures vary by measurement. A pre-release DeepInfra endpoint reportedly reached about 630 output tokens per second on 100,000-token prompts. Across six providers, Artificial Analysis gives a median of 307 tokens per second for prompts of that length, with a reported range from 106 to 534. The reported median time to the first token is 0.89 seconds, but the source says reasoning can delay the first answer token to about 11 seconds; a 500-token response took about 13.6 seconds in its account.

Where the $0.25 Task Fits

The price and speed figures make Super VL relevant to teams running high-volume document and agent workloads, particularly where they can control the model endpoint or use NVIDIA hardware. Its commercial-use license and option to run released weights on their own systems may also matter to organizations with deployment or legal requirements. Those are practical advantages, but they do not make the model the strongest value on the benchmark numbers supplied.

The source report compares Super VL’s 26.5 score and $0.25 task estimate with several alternatives. It lists MiMo-V2.6-Pro at 46 for $0.13, MiMo-V2.6-Flash at 38 for $0.06, and Inkling Small at 26 for $0.09. DeepSeek V4.1 Flash is listed at 39 for $0.27, while GLM-5.3-Flash scores 42 for $0.28. These are comparisons reported by the source, not a guarantee that every model will cost or perform the same on a buyer’s workload.

Super VL’s case is consequently more specific than “most intelligence for the money.” It may appeal when latency, throughput, licensing or self-hosting outweigh raw benchmark score. For interactive use, the reported reasoning delay could be a drawback; for background processing, a longer wait may matter less. Teams should test their own endpoint and prompts rather than treating one provider’s peak throughput as a general service level.

How NVIDIA’s Score Is Ranked

The report says Artificial Analysis places Super VL at number one among 67 open-weight models in its 40-billion-to-150-billion-parameter size class. That ranking is limited to that class: across open-weight models overall, the report puts it at about 18th. The distinction matters because a class-leading result does not mean the model leads all open-weight systems on intelligence or cost.

The report also describes its US-model comparison as license-dependent. It calls Super VL the highest-scoring permissively licensed US open-weight option, while noting that Meta’s Muse Spark Open scored 39 but uses a license with commercial-use restrictions. That characterization comes from the source report; prospective users should review the relevant license terms for their own use.

On selected task measures, the source reports gains over Nemotron 3 Super. GDPval-AA, which evaluates economically useful work, is listed at 1,162 Elo versus 490 for the predecessor. AutomationBench-AA is reported at 14.8%, compared with 3.0% for Nemotron 3 Ultra. The source also says the hallucination rate on AA-Omniscience fell from 87% to 36% relative to Nemotron 3 Super. These are benchmark results, not proof of equivalent performance in every application.

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

— Thorsten Meyer, ThorstenMeyerAI.com

What the Benchmarks Leave Open

The figures available in the source report do not establish how Super VL will perform on a particular company’s prompts, tools, or data. Artificial Analysis’s $0.25 task cost is a benchmark estimate, not a fixed charge for every real-world job. Actual bills will depend on input and output volume, caching, provider rates, and the way a workflow uses the model.

Likewise, the 630-token-per-second result came from one pre-release endpoint, while the reported cross-provider median was lower. The source does not provide enough detail here to determine whether those figures used identical serving conditions or will persist after wider availability. It also gives only selected task results; the cited hallucination rate and automation score do not settle reliability across all domains. License terms and deployment requirements should be checked directly before commercial use.

Test Costs on Real Workloads

For organizations considering deployment, the next step is to compare their own tasks, costs and response times across providers and competing models. Testing should include long-context input, cache behavior, reasoning delays, accuracy on the relevant work, and the operational cost of self-hosting versus an API. The source report advises measuring the selected endpoint because provider speed varies widely.

Artificial Analysis’s benchmark page and provider listings are the immediate reference points for updated scores and service measurements. The source material does not announce a later benchmark milestone or a company deployment schedule, so further changes to scores, availability, pricing and provider performance remain to be confirmed.

Key Questions

What does “$0.25 a task” mean?

It is Artificial Analysis’s estimated cost for one Intelligence Index task, not a universal fee for every request. Real costs depend on token use, caching, provider pricing and task design.

How capable is Nemotron 3.5 Super VL?

Artificial Analysis gives it an Intelligence Index score of 26.5. The source report says it leads its 40-billion-to-150-billion-parameter open-weight size class, but ranks about 18th among open-weight models overall.

Is 630 tokens per second the typical speed?

No. That figure was reported for a pre-release DeepInfra endpoint with 100,000-token prompts. The reported median across six providers for that prompt length was 307 tokens per second, and provider results varied.

Can businesses use the model commercially?

The source report says the model is released under OpenMDW-1.1, which allows commercial use. Users should review the license itself and confirm that its terms meet their specific deployment requirements.

Why choose it over a higher-scoring model?

Potential reasons include throughput, latency, licensing and the option to run released weights on NVIDIA hardware. The benchmark comparisons in the source show that several alternatives score higher or cost less, so the right choice depends on a team’s workload and operating constraints.

Source: ThorstenMeyerAI.com