How Nemotron 3.5 Super VL Delivers A Score Of 26

🔍 Read the full analysis: How Nemotron 3.5 Super VL Delivers A Score Of 26 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 rated it 26.5 on its Intelligence Index. The score is double the reported 13 for its predecessor, Nemotron 3 Super, and the highest among permissively licensed US open-weight models in its stated size class; it does not, however, make Super VL the strongest or cheapest model overall.

The model has 120 billion total parameters, with 12 billion active in its mixture-of-experts design, and a reported one-million-token context window. NVIDIA released weights in NVFP4, a four-bit format. The source says the OpenMDW-1.1 license allows commercial use. Artificial Analysis places the model first among 67 open-weight models in the 40-billion-to-150-billion-parameter class, but around 18th when open-weight models of all sizes are counted.

Artificial Analysis lists API prices of $0.20 per million input tokens and $0.575 per million output tokens, with cached input at $0.055 per million tokens. It estimates a cost of $0.25 per Intelligence Index task, compared with $1.64 for Nemotron 3 Super. The source attributes much of the lower task cost to reduced output volume and inexpensive cached input; it says input accounts for about 89% of the task bill in the reported breakdown.

Performance figures depend on the measurement. A pre-release DeepInfra endpoint produced about 630 output tokens per second on prompts of 100,000 tokens. Across six providers, the model page reports a median of 307 tokens per second for 100,000-token prompts, with a range from 106 to 534. Median time to the first token is reported at 0.89 seconds, while the source says reasoning can delay the first answer token to about 11 seconds.

Speed, Cost and Capability Trade-offs

For teams choosing a model, the score is only one part of the decision. Super VL’s reported speed, lower cost than its predecessor, commercial-use license and option to run released weights on NVIDIA hardware may suit organizations prioritizing deployment control or throughput. These advantages do not establish that it offers the best intelligence per dollar.

The source compares Super VL’s $0.25 task cost with models that score higher at similar or lower prices. It reports Xiaomi’s MiMo-V2.6-Pro at 46 for $0.13 per task, MiMo-V2.6-Flash at 38 for $0.06, DeepSeek V4.1 Flash at 39 for $0.27, and GLM-5.3-Flash at 42 for $0.28. Inkling Small is listed at the same score of 26 for $0.09. These comparisons are figures from the source’s Artificial Analysis-based account, not a guarantee that every model will perform equally well on a particular company’s workload.

That distinction matters operationally: a model that is quick in a benchmark may still take noticeable time to reason before responding, and an API’s provider-specific speed can differ sharply from a single endpoint result. Teams need to test their own prompts, costs and latency requirements before treating the headline score or speed figure as a production forecast.

How the Score Is Framed

The 26.5 Intelligence Index score is reported by Artificial Analysis, not presented here as a universal measure of model quality. The source’s class ranking covers open-weight models with 40 billion to 150 billion total parameters. Its broader ranking changes when models outside that size range are included, so the “top” claim applies only to the specified comparison group and licensing category.

The source describes a marked change from Nemotron 3 Super: the predecessor scored 13 on the index and cost a reported $1.64 per task. It also says the new model uses far fewer output tokens per index task—about 46,000 compared with 114,000. Reported task results include a rise in GDPval-AA, a benchmark for economically useful work, from 490 Elo to 1,162 Elo, and a rise in AutomationBench-AA to 14.8% from 3.0%. Those benchmark results indicate progress on the tested tasks, not a general guarantee of reliable automation.

The source also says the AA-Omniscience hallucination rate fell from 87% for the predecessor to 36% for Super VL. That remains a reported benchmark result, not evidence that hallucinations have been eliminated. Separately, the source notes that Meta’s Muse Spark Open scored 39 but uses a license with commercial-use restrictions, which makes license terms a separate consideration from model score.

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

— Thorsten Meyer, ThorstenMeyerAI.com

Provider Results and Real-World Limits

The reported 630-token-per-second result came from one pre-release DeepInfra endpoint. The source gives a substantially lower median across six providers, but does not identify in the supplied material how each provider’s hardware, configuration or serving conditions affected the spread. Organizations should not assume the fastest result will match their chosen endpoint.

The available figures also leave open how well benchmark performance transfers to specific production workloads. The source reports strong gains on selected knowledge-work and automation tests, alongside a 36% hallucination rate on AA-Omniscience. The supplied material does not provide enough detail to assess all test methods, confidence ranges, or the model’s performance across different languages and image-based tasks. Its one-million-token context specification likewise does not establish that every long-context task will be accurate or economical.

Licensing and deployment details may require review by each user. The source says commercial use is allowed under OpenMDW-1.1 and that the weights can be run on NVIDIA hardware, but it does not give a complete account of hardware requirements, operational costs or license conditions. The headline ranking also depends on the comparison set: it is not a claim that Super VL leads all open-weight models.

Test Before Production Deployment

The immediate next step for prospective users is to benchmark the model on their own workloads. That means checking the provider and hardware they intend to use, measuring response latency as well as throughput, and tracking input, cache and output token costs separately. The source specifically advises measuring the endpoint a team plans to use rather than relying on the single-endpoint speed result.

Teams evaluating self-hosting will also need to verify that the released NVFP4 weights fit their hardware and that the license meets their legal and commercial requirements. The supplied source does not give a date for further model updates or identify a next benchmark milestone. Performance comparisons may change as providers add support or new models enter the evaluated sets.

Key Questions

What does Nemotron 3.5 Super VL’s score of 26 mean?

Artificial Analysis reports an Intelligence Index score of 26.5. The source says this is double Nemotron 3 Super’s score of 13 and ranks first among 67 open-weight models in the 40B-to-150B size class. It is not the top score among all open-weight models.

Is Nemotron 3.5 Super VL the best-value open model?

Not according to the price and score comparisons in the source. Several listed alternatives score higher for similar or lower reported cost per index task, while Inkling Small is listed at the same score for less. The best choice depends on workload, provider, license and deployment needs.

How fast is Nemotron 3.5 Super VL?

A pre-release DeepInfra endpoint reportedly produced about 630 output tokens per second on 100,000-token prompts. The reported median across six providers for that prompt length is 307 tokens per second, with results ranging from 106 to 534.

Can businesses use the model commercially?

The source says the model is released under the OpenMDW-1.1 license, which allows commercial use. Organizations should review the license itself and confirm that its conditions meet their intended use.

What is not yet clear about the model?

The supplied material does not establish how benchmark results will translate to each organization’s tasks, how much self-hosting will cost, or why provider speeds vary. It also does not show that the reported hallucination rate represents all use cases.

Source: ThorstenMeyerAI.com