Nemotron 3.5 Super VL Brings Together Fast AI And Low Task Costs

🔍 Read the full analysis: Nemotron 3.5 Super VL Brings Together Fast AI And Low Task Costs on ThorstenMeyerAI.com

Before you orderOffer from Amazon

Get the latest gadgets 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, and Artificial Analysis published evaluations the same day, reporting a 26 Intelligence Index score and a $0.25 cost for one index task. The 120-billion-parameter open-weight model is designed to run quickly, but comparisons in the evaluation show that several alternatives deliver higher scores for similar or lower task costs.

Artificial Analysis lists the model at 26.5 points in its detailed data, while the launch write-up rounds that to 26. It ranks first among 67 open-weight models in the 40-billion to 150-billion parameter class, according to the supplied evaluation. Counting open-weight models across all sizes, the article places it around 18th. The model has 120 billion total parameters and 12 billion active parameters in a mixture-of-experts design, accepts text and images, and has a listed context window of one million tokens.

The evaluation reports a price of $0.20 per million input tokens and $0.575 per million output tokens through the listed API, with cached input at $0.055 per million tokens. Artificial Analysis puts the average index-task cost at $0.25, compared with $1.64 for Nemotron 3 Super. The supplied account attributes the lower task cost partly to fewer generated output tokens and inexpensive cached input. It says roughly 89% of the $0.25 task bill is input, including cache writes, while output accounts for less than three cents.

Speed varies by measurement. A pre-release DeepInfra endpoint serving NVFP4 weights produced about 630 output tokens per second with prompts of 100,000 tokens. Across six providers, the reported median was 196 tokens per second on the standard prompt and 307 with 100,000-token prompts; the measured range ran from 106 to 534. Median time to first token was 0.89 seconds, but the account says reasoning delayed the first answer token by about 11 seconds and a 500-token answer took about 13.6 seconds end to end.

Speed Comes With a Value Trade-Off

For teams running large document or repository workloads, the reported low cost per index task and discounted cached input may matter more than the headline token prices. NVIDIA also released the model weights in NVFP4, and the stated OpenMDW-1.1 license allows commercial use. The article says teams can run the model on their own NVIDIA hardware, giving operators an option to manage deployment themselves.

Those benefits do not make it the strongest model for the money by the reported Intelligence Index measure. Xiaomi’s MiMo-V2.6-Pro scores 46 at a reported $0.13 per task, while MiMo-V2.6-Flash scores 38 at $0.06. DeepSeek V4.1 Flash scores 39 at $0.27, and GLM-5.3-Flash scores 42 at $0.28. Inkling Small matches the rounded score of 26 at $0.09. These figures make task cost and capability separate questions: Super VL’s case depends on speed, deployment control and license fit, as well as its score.

The release account also reports gains over the previous Nemotron model on selected measures. GDPval-AA, an evaluation of economically useful tasks, rose to 1,162 Elo from 490; AutomationBench-AA reached 14.8%, compared with 3.0% for Nemotron 3 Ultra. The account describes the automation result as a large relative gain, while noting that the absolute score remains low. These benchmark results may help identify where to evaluate the model, but they do not establish how it will perform on any particular company’s workflows.

How the Model Ranks

The headline ranking needs its size-class qualifier. Artificial Analysis compares open-weight models in the same parameter range; Super VL’s reported first-place position applies to the 40-billion to 150-billion total-parameter group, not all open-weight models. The supplied article estimates its overall open-weight position at about 18th. The distinction matters because a class-leading result is not the same as a top overall score.

The release follows Nemotron 3 Super, whose reported Intelligence Index score was 13. The new model’s score is roughly double that predecessor’s, while its reported index-task cost falls from $1.64 to $0.25. Those comparisons describe the specific Artificial Analysis index and task-cost methodology. They are not a guarantee that every user will see the same cost reduction or quality change on a different workload.

License comparisons also depend on the user’s requirements. The source describes Super VL as the top-scoring US open-weight option among models with permissive licenses. Meta’s Muse Spark Open, released on the same day, scored 39 but uses the Muse Model License, which the article says permits commercial use with restrictions. Teams should review the relevant license terms for their intended use before deciding which model qualifies for them.

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

— Thorsten Meyer, in the October 16, 2026, ThorstenMeyerAI.com article

Provider Results and Workloads Vary

The speed results do not establish what every deployment will achieve. The 630-token-per-second figure came from one pre-release endpoint, while provider-wide measurements ranged from 106 to 534 tokens per second. The article advises measuring the specific endpoint a team plans to use. It does not provide a comparable end-to-end latency breakdown for each provider.

The supplied material reports a hallucination rate of 36% on AA-Omniscience, down from 87% for Nemotron 3 Super, but it does not give further methodological detail or explain how that measure translates to individual use cases. The source also stops before providing the full evaluation profile. Results on particular image, reasoning, or production tasks, and the hardware requirements and operating costs for self-hosting, are not established by the figures provided.

Test the Intended Deployment

Teams considering the model can compare it on their own tasks, using the intended provider, prompt lengths, caching pattern and response requirements. That evaluation should track both quality and end-to-end latency, since the reported first-token delay under reasoning differs from raw output throughput. Operators can also calculate costs using their expected input and cache volumes, rather than relying on output-token prices alone.

Further provider measurements and fuller benchmark results could clarify how stable the speed and quality findings are across deployments. Until then, the available evidence supports a specific conclusion: Nemotron 3.5 Super VL offers a commercially usable open-weight option with strong results in its size class and promising reported economics, while its overall value depends on whether speed, license terms and self-hosting matter more to a team than a higher index score.

Key Questions

When was Nemotron 3.5 Super VL released?

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

What does the score of 26 mean?

It is the rounded score reported for the Artificial Analysis Intelligence Index; the detailed figure in the source is 26.5. The source places the model first among 67 open-weight models in the 40-billion to 150-billion parameter class, and around 18th across open-weight models overall.

How fast is the model?

One pre-release DeepInfra endpoint reached about 630 output tokens per second with 100,000-token prompts. The median across six providers was 307 tokens per second at that prompt length, with results varying by provider.

How much does an index task cost?

Artificial Analysis reports an average cost of $0.25 per index task. Actual costs depend on the workload and provider; the source says input and cache writes make up most of the measured task bill.

Can businesses use the model commercially?

The source identifies its license as OpenMDW-1.1 and says commercial use is allowed. Businesses should check the license terms against their planned deployment.

Source: ThorstenMeyerAI.com