How Nemotron 3.5 Super VL Balances Cost, Speed, And Score

🔍 Read the full analysis: How Nemotron 3.5 Super VL Balances Cost, Speed, And Score 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 a score of 26 on its Intelligence Index, double the 13 recorded for Nemotron 3 Super. The new model also costs less per benchmark task than its predecessor, but the score and pricing do not make it the strongest value on raw intelligence: several models in the source comparison score higher for a similar or lower task cost.

Artificial Analysis places Super VL first among 67 open-weight models in its 40-billion-to-150-billion-parameter size class. Its reported score is 26.5 in the specification table, rounded to 26 in the article text. Across all open-weight models counted once, the source estimates its position at about 18th. NVIDIA’s model has 120 billion total parameters and 12 billion active in a mixture-of-experts design, a stated one-million-token context window, and accepts text and images while producing text. Its released weights are in NVFP4, a four-bit format.

The listed API rates are $0.20 per million input tokens and $0.575 per million output tokens, with cached input at $0.055 per million tokens. Artificial Analysis estimates a full index task at $0.25, compared with $1.64 for Nemotron 3 Super. That is an approximately 85% lower task cost alongside twice the predecessor’s score, according to the supplied figures. The source says the newer model generated about 46,000 output tokens per index task, against 114,000 for its predecessor, and that input accounted for about 89% of the $0.25 total.

Speed measurements depend on where the model runs. A pre-release DeepInfra endpoint using NVFP4 weights produced about 630 output tokens per second on 100,000-token prompts. Artificial Analysis’ median across six providers was 307 tokens per second on those prompts and 196 on its standard prompt. The reported 100,000-token speed ranged from 106 tokens per second at the fifth percentile to 534 at the 95th. Median time to first token was 0.89 seconds, while the source says reasoning delayed the first answer token to about 11 seconds; a 500-token answer took about 13.6 seconds end to end.

Where the Cost Advantage Holds

The release matters most to teams choosing between inference cost, response speed and control over deployment. At the stated rates, a workload dominated by repeatedly supplied or cached documents may have a different bill from one dominated by generated output: the source’s index-task breakdown attributes most of its estimated cost to input, including cache writes. Teams should compare their own input mix and caching behavior with the benchmark task rather than treating the per-token rates as a complete cost forecast.

Super VL’s reported speed could suit background agents and pipeline work, where throughput and self-hosting matter more than an immediate conversational reply. But the gap between one endpoint’s 630 tokens per second and the six-provider median of 307 shows why a benchmark peak is not a dependable production estimate. The source also reports a roughly 11-second reasoning period before the first answer token, a potential drawback in interactive applications.

Its score places the trade-off in perspective. 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. Those figures are Artificial Analysis comparisons, not a guarantee that each model will perform identically on a reader’s workload. Still, they show that Super VL’s case is not simply more intelligence for the money; it rests on speed, licensing, model provenance and the option to run weights on NVIDIA hardware.

The Benchmarks Behind the Release

The Artificial Analysis Intelligence Index is the central comparison cited in the source. Its size-class ranking is narrower than an all-model ranking: first among models in the stated 40B–150B parameter band does not mean first among every open-weight model. The source reports the score as 26 in its headline comparison and 26.5 in its specification table, so the rounded figure is used here where a single number is needed.

Some benchmark categories show larger gains over Nemotron 3 Super than the aggregate index alone conveys. Artificial Analysis figures cited by the source put Super VL at 1,162 Elo on GDPval-AA, compared with 490 for the predecessor, and 999 on AA-Briefcase, compared with 876 for Nemotron 3 Ultra. On AutomationBench-AA, the reported result is 14.8%, versus 3.0% for Nemotron 3 Ultra. The source describes that as about five times higher while also noting the new model’s absolute score remains low.

On AA-Omniscience, the source reports a hallucination rate of 36%, down from 87% for Nemotron 3 Super. That is a substantial change in the cited evaluation, but it does not establish that hallucinations are eliminated or predict error rates across all uses. The source also distinguishes licensing claims: it describes Super VL’s OpenMDW-1.1 license as allowing commercial use, while saying Meta’s same-day Muse Spark Open uses a license with commercial-use restrictions.

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

— Thorsten Meyer, ThorstenMeyerAI.com

Production Results Still Vary

The cited benchmarks and prices do not establish how Super VL will perform on a particular company’s prompts, hardware or provider endpoint. The 630-token-per-second figure came from one pre-release endpoint; the source’s cross-provider median was lower, and the reported range was wide. No single figure in the source settles the speed users should expect in production.

The source does not provide a workload-by-workload cost forecast, independent replication of all benchmark results, or details sufficient to compare self-hosting expenses with API use. It also does not spell out every practical or legal condition attached to the named licenses. The reported hallucination, automation and knowledge-work results are specific to cited evaluations; performance outside those tests remains uncertain.

There is also a limit to what the supplied material establishes about current availability: it says the model was released and lists six providers, but does not identify all providers or confirm that each offers the same weight format, latency or pricing. The reported comparisons with other models should be read as the source’s benchmark snapshot, not as a universal ranking for every task.

Test the Intended Deployment

For teams considering adoption, the next useful step is to test the specific provider or NVIDIA hardware configuration they intend to use. A representative evaluation should track time to first answer, full response time, input and output token volumes, cache behavior, task accuracy and cost. The source’s wide provider spread makes local measurement more informative than relying on the fastest reported endpoint alone.

Readers should also check the current model documentation and license terms before deployment, particularly if commercial use or self-hosting is a requirement. The source gives a release date and benchmark snapshot but does not state a later update or a next benchmark milestone. Further provider measurements and independent testing would clarify whether the reported price and speed advantages persist across real production workloads.

Key Questions

What score did Nemotron 3.5 Super VL receive?

Artificial Analysis reported an Intelligence Index score of 26; the source’s specification table gives 26.5. It ranks first among 67 open-weight models in the cited 40B–150B size class, while the source estimates its overall open-weight position at about 18th.

How much does an Artificial Analysis index task cost?

The source estimates $0.25 per index task, compared with $1.64 for Nemotron 3 Super. That is a benchmark estimate; actual bills depend on the workload, provider, token volumes and caching.

How fast is Super VL?

A pre-release DeepInfra endpoint reached about 630 output tokens per second on 100,000-token prompts. Across six providers, the reported median at that prompt length was 307 tokens per second, with measurements ranging from 106 to 534 at the fifth and 95th percentiles.

Does the reported score make it the best-value model?

Not on the source’s raw score-and-task-cost comparisons. It lists several models with higher scores for similar or lower estimated task costs, and Inkling Small with the same score of 26 at a lower estimated cost. Super VL may appeal for other reasons, including speed, its stated commercial-use license and self-hosting.

What remains unknown about using it in production?

Results on a reader’s workload, real deployment costs and speed on a chosen endpoint are not established by the cited figures. The source also does not provide full details for every provider or a detailed account of license conditions, so teams should verify current documentation and test their intended setup.

Source: ThorstenMeyerAI.com