Three Numbers To Know About Nemotron 3.5 Super VL

🔍 Read the full analysis: Three Numbers To Know About Nemotron 3.5 Super VL on ThorstenMeyerAI.com

Before you orderOffer from Amazon

Get hardware and tech essentials delivered free with Prime

• Fast, free delivery on millions of items
• Prime Video, Amazon Music and more included
• Member-only deals all year

Start your free Prime trial →Free trial for eligible customers · Cancel anytime

As an affiliate, we earn on qualifying purchases.Hide

NVIDIA released Nemotron 3.5 Super VL on October 15, 2026, and early Artificial Analysis results put it at 26 on the Intelligence Index, with a reported $0.25 cost per index task. The figures make the model notable for teams weighing speed, cost and commercial-use licensing, but the headline speed of roughly 630 output tokens per second came from one pre-release provider endpoint, not the median across services.

The three numbers frame different parts of the release. The score of 26 is the model’s Artificial Analysis Intelligence Index result; the source reports that this is twice Nemotron 3 Super’s score of 13. It ranks first among 67 open-weight models in its 40-billion-to-150-billion-parameter size class, according to the source. That is a class-specific ranking, not a claim that it leads all open-weight models: the source places it about 18th when models of all sizes are counted.

The $0.25 per task figure is Artificial Analysis’ cost estimate for one Intelligence Index task, not a general price for every user request. The source lists API rates of $0.20 per million input tokens and $0.575 per million output tokens, with cached input at $0.055 per million. It says input accounts for about 89% of the benchmark-task cost, including cache writes, while output is less than three cents. Actual bills depend on a workload’s token use, caching and provider.

The speed figure needs its measurement conditions. On a pre-release DeepInfra endpoint serving NVFP4 weights, the model produced about 630 output tokens per second with 100,000-token prompts. Across six providers, the source reports a median of 196 tokens per second on its standard prompt and 307 with 100,000-token prompts. The measured provider range was 106 to 534 tokens per second. Median time to first token was 0.89 seconds, but the source estimates about 11 seconds before the first answer token when the model reasons before responding.

Three Metrics, Different Trade-Offs

Together, the figures describe a model that may appeal to operators who value fast processing, controlled infrastructure and a permissive commercial-use license, rather than maximum benchmark performance per dollar. NVIDIA’s release uses a mixture-of-experts design with 120 billion total parameters and 12 billion active, a one-million-token context window, and image-and-text input with text output. Its weights were released in NVFP4, a four-bit format, and the source says users can run the model on their own NVIDIA hardware.

That option to self-host may matter to organizations with deployment, data-control or licensing requirements. But the source’s comparisons show that a score of 26 is not the strongest value on raw index performance alone. It lists models with higher scores at similar or lower estimated task costs, including MiMo-V2.6-Pro at 46 for $0.13 and Inkling Small at 26 for $0.09. These comparisons use the source’s reported benchmark and cost figures; they do not establish that models are interchangeable for every task or deployment.

The result is a practical distinction: the model’s case rests on more than its score. Speed can help in high-volume pipelines, while latency and reasoning time affect interactive use. License terms and the ability to run weights locally may also influence a procurement decision. Those advantages should be evaluated against measured performance on a team’s own prompts, provider and hardware.

How the Benchmark Fits

The model arrived as an update to Nemotron 3 Super, whose Intelligence Index score the source gives as 13. The reported task cost fell from $1.64 for the predecessor to $0.25 for Super VL, while the score doubled. The source attributes much of the cost difference to fewer generated tokens: about 46,000 output tokens per index task for Super VL, compared with 114,000 for Nemotron 3 Super.

The source also reports improvement in selected evaluations. On GDPval-AA, a benchmark of economically useful tasks, Super VL scored 1,162 Elo, compared with 490 for its predecessor. Its AutomationBench-AA result was 14.8%, versus 3.0% for Nemotron 3 Ultra; the source cautions that the absolute result remains low. On AA-Omniscience, it reports a hallucination rate of 36%, down from 87% for Nemotron 3 Super. These are benchmark outcomes as reported by the source, not guarantees about performance on a specific organization’s workload.

License comparisons also need care. The source describes OpenMDW-1.1 as allowing commercial use and says Super VL leads among permissively licensed US open-weight models. It contrasts that with Meta’s Muse Spark Open, which it says scored 39 but uses a license with commercial-use restrictions. The label “open-weight” alone does not settle whether a model’s terms fit a particular use; teams should review the actual license.

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

— Thorsten Meyer, ThorstenMeyerAI.com

What the Early Figures Cannot Show

The results do not establish how quickly every deployment will run. The 630-token-per-second figure came from one pre-release endpoint, while the six-provider median was lower and the provider range was broad. Hardware, quantization, prompt length, service configuration and load can affect observed performance. The source does not provide a controlled comparison of every model on identical infrastructure and workloads.

Nor does a single benchmark-task estimate establish the cost of an organization’s production use. The $0.25 figure reflects the Artificial Analysis task profile; workloads with repeated document ingestion, different caching behavior or longer outputs may produce different bills. The source does not specify how broadly the reported hallucination improvement carries across real-world domains, or provide enough detail here to assess uncertainty in each benchmark result.

It also remains unclear from the source how the model performs across languages, specialized tasks and varying image inputs, and how long the reported provider prices and availability will remain in place. No independent production deployment results are included. Readers should treat the launch figures as reported evaluations rather than a substitute for task-specific testing.

Test Providers and Workloads

The next useful step for prospective operators is to compare the available providers using the prompts, context lengths and response formats they expect to use. Track end-to-end latency, output quality and cost per completed task, rather than relying only on tokens per second or listed token rates. Teams considering self-hosting should also test the released NVFP4 weights on their intended NVIDIA hardware and review the OpenMDW-1.1 terms for their application.

As more deployment data becomes available, comparisons can show whether the launch benchmarks hold under sustained use and whether the reported gains in knowledge work, automation and hallucination rates transfer to particular workloads. Until then, the confirmed picture is a score of 26, a benchmark-task estimate of $0.25 and a high speed reading from one endpoint, alongside substantially lower median speeds across providers.

Key Questions

What are the three numbers to know about Nemotron 3.5 Super VL?

The source highlights an Intelligence Index score of 26, an estimated $0.25 per index task, and about 630 output tokens per second on one pre-release endpoint with 100,000-token prompts. The speed figure is not the multi-provider median.

Is 630 tokens per second the typical speed?

No. The source reports a six-provider median of 196 tokens per second on its standard prompt and 307 with 100,000-token prompts. Results ranged from 106 to 534 tokens per second across providers; the 630 figure came from one pre-release DeepInfra endpoint.

Does the $0.25 figure represent every request?

No. It is Artificial Analysis’ estimated cost for one benchmark task. API rates are listed separately, and actual spending depends on input and output token counts, caching, provider and workload.

Can businesses use the model commercially?

The source says the OpenMDW-1.1 license allows commercial use. Organizations should review the license itself and confirm that its terms fit their intended deployment.

What remains unverified for production use?

The source does not establish performance across all hardware, providers, languages or task types. Production users still need to test quality, speed, latency and cost on their own workloads; AI model performance and costs can vary, and deployment carries operational and financial risks.

Source: ThorstenMeyerAI.com