Exploring Nemotron 3.5 Super VL’s Cost Per Task And Latency

🔍 Read the full analysis: Exploring Nemotron 3.5 Super VL’s Cost Per Task And Latency 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, and Artificial Analysis reported an Intelligence Index score of 26.5 and an estimated $0.25 per index task. The 120-billion-parameter open-weight model is fast on some endpoints and allows commercial use under its OpenMDW-1.1 license, but comparisons in the source data show several models scoring higher for 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. That ranking has a narrower scope than a ranking of all open-weight models: counted once each, Super VL is about 18th overall, according to the source article. The model has 120 billion total parameters, with 12 billion active under its mixture-of-experts design, and accepts text and images while producing text. Its listed context window is one million tokens.

The published API rates are $0.20 per million input tokens and $0.575 per million output tokens; cached input is listed at $0.055 per million. Artificial Analysis estimates one index task at $0.25, compared with $1.64 for Nemotron 3 Super. The source attributes the lower task bill partly to fewer output tokens: about 46,000 per index task versus 114,000 for the predecessor. It says roughly 89% of Super VL’s task cost is input, with cache writes accounting for 69% and output costing less than three cents.

Speed figures depend on where the model runs. A pre-release DeepInfra endpoint serving NVFP4 weights generated about 630 output tokens per second with 100,000-token prompts, according to Artificial Analysis. Across six providers, the model page reports a median of 196 tokens per second on standard prompts and 307 with 100,000-token prompts; the reported 5th-to-95th-percentile range is 106 to 534. Median time to first token is 0.89 seconds, but the source says reasoning takes about 11 seconds before the first answer token, and a 500-token answer takes about 13.6 seconds end to end.

The Trade-Off Behind Low Task Costs

For operators, the figures point to a specific cost profile: input usage can dominate the bill, particularly when an agent repeatedly reads a large repository or document collection. Cached input is cheaper than uncached input at the listed rates, but actual spending will still depend on prompt size, cache behavior and task design. A low per-task benchmark estimate does not by itself predict an individual deployment’s monthly bill.

Speed also depends on the serving provider. The 630-token-per-second result came from one pre-release endpoint, while the six-provider median was lower and the reported range broad. Teams considering Super VL for interactive use should account for the reported reasoning delay; it may fit background workflows differently from a chat experience where users wait for an answer.

Artificial Analysis comparisons in the source article complicate a simple value claim. MiMo-V2.6-Pro scored 46 at $0.13 per task, while MiMo-V2.6-Flash scored 38 at $0.06. DeepSeek V4.1 Flash scored 39 at $0.27, and Inkling Small scored 26 at $0.09. Those figures suggest Super VL’s case depends on factors beyond benchmark score per dollar, including its license, NVIDIA provenance, latency and the option to run its released weights on NVIDIA hardware.

How the Benchmark Rankings Compare

The Intelligence Index score is a benchmark summary, not a measure of every production task. Artificial Analysis gave Nemotron 3 Super VL a score of 26.5, rounded to 26, twice the predecessor’s score of 13. The source article describes it as the top-scoring permissively licensed US open-weight model, while noting that the category depends on licensing criteria: Meta’s Muse Spark Open scored 39 but uses a license that permits commercial use with restrictions.

The source also reports gains on selected evaluations. GDPval-AA, covering economically useful work, rose to 1,162 Elo from 490 for Nemotron 3 Super; the article says that is above Nemotron 3 Ultra’s 1,013 and Inkling Small’s 1,124. Super VL scored 999 on AA-Briefcase, compared with 876 for Ultra. AutomationBench-AA was 14.8%, versus 3.0% for Ultra, though the source characterizes the absolute result as low. Its reported AA-Omniscience hallucination rate fell to 36% from 87% for the predecessor. These are benchmark results, not guarantees of performance on a particular company’s workload.

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

— Thorsten Meyer, ThorstenMeyerAI.com

Provider Results and Real-World Fit

The source figures do not establish what speed or cost a given team will see in production. Provider, hardware, prompt length, caching and workload can change results; the reported 630 tokens per second was measured on one pre-release endpoint. The source material does not provide a full methodology for reproducing every cost estimate or benchmark result.

It is also unclear how the model will perform across specific image tasks, languages, safety requirements or production workloads, since the supplied material gives no detailed results for those areas. The benchmark comparisons do not settle which licensing terms best fit a particular organization. Super VL’s license allows commercial use according to the source, but teams still need to assess their own requirements and the terms of the actual service or deployment they choose.

Check the Intended Deployment

Teams evaluating Super VL can compare its released NVFP4 weights and API options against their own prompts, image inputs and response requirements. A representative pilot should record latency, throughput and input-token use on the intended provider, including the effect of caching. The source article does not identify a future benchmark update or a next release date, so the next useful evidence will come from measurements on the endpoints and workloads teams plan to use.

Key Questions

When did NVIDIA release Nemotron 3.5 Super VL?

NVIDIA released the model on October 15, 2026, according to the source article. Artificial Analysis published benchmark figures the same day.

How much does one benchmark task cost?

Artificial Analysis estimated $0.25 per Intelligence Index task. The source says this is a benchmark cost estimate; an individual deployment’s bill will depend on its input volume, caching and workload.

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 from the 5th to 95th percentile.

Is Nemotron 3.5 Super VL the top open-weight model overall?

No. The source says it ranked first among 67 open-weight models in its 40-billion-to-150-billion-parameter size class, but about 18th among open-weight models overall. It also lists several models with higher Intelligence Index scores.

Can companies use the model commercially?

The source identifies its license as OpenMDW-1.1 and says commercial use is allowed. Organizations should check the applicable license and provider terms for their intended deployment.

Source: ThorstenMeyerAI.com