The Price And Pace Of Nemotron 3.5 Super VL

🔍 Read the full analysis: The Price And Pace Of Nemotron 3.5 Super VL on ThorstenMeyerAI.com

Buying for a business?Offer from Amazon

Get business pricing on office and shipping supplies

• 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, roughly double its predecessor’s score. The model combines low reported cost per benchmark task and high throughput on one test endpoint with open weights and a license allowing commercial use, but several alternatives scored higher for similar or lower reported costs.

Artificial Analysis lists the model at $0.20 per million input tokens and $0.575 per million output tokens, with cached input priced at $0.055 per million tokens. It estimates a full Intelligence Index task costs $0.25, compared with $1.64 for Nemotron 3 Super. Artificial Analysis attributes much of the reduction to fewer output tokens: about 46,000 per index task, against 114,000 for the predecessor. Its cost breakdown puts about 89% of the new model’s task cost on input, including cache writes.

The model has 120 billion total parameters and 12 billion active parameters in a mixture-of-experts design, and a reported context window of one million tokens. NVIDIA released weights in NVFP4, a four-bit format. The source says its OpenMDW-1.1 license permits commercial use. Artificial Analysis counts six providers, whose speed results vary substantially.

Artificial Analysis measured about 630 output tokens per second on a pre-release DeepInfra endpoint using 100,000-token prompts. That was the fastest result it had measured at that prompt length among models scoring at least 20, according to its report. Across six providers, however, the model page gives a median of 307 tokens per second on 100,000-token prompts, with a range from 106 at the fifth percentile to 534 at the 95th percentile. The median first-token latency is 0.89 seconds; the source estimates about 11 seconds to the first answer token when reasoning is included.

The Trade-Off Behind the Score

The release offers teams a model with commercial-use terms and downloadable weights, alongside a reported cost per benchmark task far below that of its predecessor. Those attributes may matter to organizations that want to run inference on their own NVIDIA hardware or control where workloads are handled. The source does not provide a full cost comparison for self-hosting, so the $0.25 figure should be read as the benchmark’s reported task cost, not a complete estimate of operating the model locally.

For buyers choosing mainly by score per dollar, the comparisons in the source complicate the pitch. Xiaomi’s MiMo-V2.6-Pro scores 46 for a reported $0.13 per task, while MiMo-V2.6-Flash scores 38 for $0.06. Inkling Small matches Super VL’s score of 26 at a reported $0.09. These figures come from the cited Artificial Analysis comparisons; benchmark results do not establish how a model will perform on every workload.

The performance profile also differs by task. Artificial Analysis reports a large gain over Nemotron 3 Super on GDPval-AA, a test of economically useful work, and an increase on SaaS automation. But its reported AutomationBench-AA result is still 14.8%. Teams need to assess their own tasks, provider speed and input patterns before treating benchmark price or throughput as a production estimate.

What the Benchmarks Compare

Artificial Analysis reports a score of 26.5, rounded to 26, on its Intelligence Index. It places Super VL first among 67 open-weight models in its 40 billion to 150 billion total-parameter size class. That class result is narrower than a claim that it leads all open-weight models: the source places it around 18th when counting open-weight models overall.

The source describes the model as the top-scoring permissively licensed US open-weight model. That distinction matters because the comparison excludes models whose terms impose more restrictions. Meta’s Muse Spark Open, released on the same day according to the source, scores 39 but uses a license that permits commercial use with restrictions. “Top” therefore depends on the license and geographic criteria being applied.

Artificial Analysis also reports a decline in the predecessor’s hallucination rate on its AA-Omniscience evaluation, from 87% to 36% for Super VL. This is a result on a specific evaluation, not a guarantee of accuracy in deployed systems. The source’s other cited results include 1,162 Elo on GDPval-AA, compared with 490 for Nemotron 3 Super, and 999 Elo on AA-Briefcase, compared with 876 for Nemotron 3 Ultra.

Questions for Production Use

The reported 630-token-per-second result does not describe every provider. Artificial Analysis’ six-provider results range from 106 to 534 tokens per second at the reported percentiles, and provider selection can materially change observed speed. The source does not establish which endpoint a particular team will receive or what performance it will see under its own traffic.

It is also unclear how the model will perform across a specific company’s documents, tools and failure conditions. A lower hallucination rate on AA-Omniscience does not settle the risk of incorrect answers in other settings. The supplied material does not give detailed per-task error rates for all listed comparisons, nor does it establish total self-hosting costs or hardware requirements for a given production workload.

Measure Before Deployment

Teams considering Super VL can compare its reported benchmark results with their own task set, then measure latency, throughput and token use on the provider or hardware they expect to run. Long-context applications should track input and cache costs, since the source’s breakdown attributes most of the benchmark task bill to input.

Further provider measurements and workload-specific evaluations will clarify whether the model’s speed, license and deployment options outweigh its lower index score relative to some alternatives. The release material does not specify a next benchmark milestone or a planned update date.

Key Questions

When did NVIDIA release Nemotron 3.5 Super VL?

NVIDIA released it on October 15, 2026, according to the supplied source material.

What does the Intelligence Index score measure here?

Artificial Analysis reports a score of 26.5, often rounded to 26. The supplied material places the model first among 67 open-weight models in its 40 billion to 150 billion parameter size class, and around 18th among open-weight models overall.

How much does an Intelligence Index task cost?

Artificial Analysis estimates $0.25 per task. Its listed API prices are $0.20 per million input tokens and $0.575 per million output tokens, with cached input at $0.055 per million tokens. Actual costs depend on usage and provider.

Is the 630-token-per-second speed typical?

It was measured on one pre-release DeepInfra endpoint with 100,000-token prompts. Artificial Analysis reports a six-provider median of 307 output tokens per second at that prompt length, with a fifth-to-95th-percentile range of 106 to 534.

Does its lower hallucination result guarantee more reliable answers?

No such guarantee follows from the supplied figures. Artificial Analysis reports a fall from 87% to 36% on AA-Omniscience compared with Nemotron 3 Super. That evaluation result does not establish accuracy for every task or deployment.

Source: ThorstenMeyerAI.com