Could Nemotron 3.5 Super VL Make AI Tasks More Affordable?

🔍 Read the full analysis: Could Nemotron 3.5 Super VL Make AI Tasks More Affordable? on ThorstenMeyerAI.com

Before you orderOffer from Amazon

Get the little things that make your day 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, with Artificial Analysis reporting an Intelligence Index score of 26 and an estimated cost of $0.25 per index task. The model is markedly cheaper per task than its predecessor and can run on NVIDIA hardware, but comparisons cited by the evaluator show that several competing models score higher for similar or lower costs.

Artificial Analysis assigns Super VL a score of 26.5, rounded to 26 in the index, and places it first among 67 open-weight models in its 40-billion-to-150-billion-parameter size class. It has 120 billion total parameters, with 12 billion active in its mixture-of-experts design, and a reported one-million-token context window. The ranking is specific to that size class: the source places it about 18th among open-weight models overall.

The model is released in NVFP4, a four-bit format, and uses the OpenMDW-1.1 license, which the source describes as permitting commercial use. 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 an average $0.25 per benchmark task; that figure reflects its evaluation workload, not a guaranteed price for every real-world request.

Speed varies by measurement. A pre-release DeepInfra endpoint returned about 630 output tokens per second on prompts containing 100,000 tokens. Across six providers, however, Artificial Analysis reports a median of 196 tokens per second on its standard prompt and 307 on 100,000-token prompts, with provider results ranging from 106 to 534 tokens per second. The reported median time to first token is 0.89 seconds, while the source says reasoning can delay the first answer token to about 11 seconds.

Cost Savings Versus Model Quality

For teams running repeated agent or document-processing workloads, the appeal is a combination of lower benchmark cost, fast generation and self-hosting. Artificial Analysis puts Super VL at $0.25 per index task, compared with $1.64 for Nemotron 3 Super. The source attributes the reduction partly to lower output volume: about 46,000 generated tokens per index task versus 114,000 for the older model. It also says roughly 89% of Super VL’s estimated task cost is input, including cache writes, so workloads that repeatedly process long prompts may have very different bills from those suggested by output-token rates alone.

That cost improvement does not make Super VL the strongest option on the source’s comparisons. Xiaomi’s MiMo-V2.6-Pro is listed at 46 points and $0.13 per task; MiMo-V2.6-Flash at 38 points and $0.06; and Inkling Small at 26 points and $0.09. Those figures make raw benchmark score per dollar a separate question from speed, deployment control or licensing. Results may matter differently for operators who prioritize running weights on their own infrastructure or need a permissive license.

The source also reports gains over the predecessor in selected evaluations: GDPval-AA, which tests economically useful work, rises from 490 to 1,162 Elo, while AutomationBench-AA reaches 14.8%, compared with Nemotron 3 Ultra’s 3.0%. It reports a hallucination rate of 36% on AA-Omniscience, down from 87% for Nemotron 3 Super. These are benchmark-specific results, not evidence that the model will perform equally well across every deployment.

How the Release Is Evaluated

The release follows NVIDIA’s Nemotron 3 Super. In Artificial Analysis’ reported figures, the earlier model scored 13 on the Intelligence Index and cost $1.64 per index task; Super VL scores 26 and is estimated at $0.25. That is a substantial change on the same named index, though benchmark scores and task-cost estimates should not be treated as direct predictions of an organization’s production quality or total operating expense.

The size-class ranking also needs its label. Artificial Analysis compares Super VL with open-weight models containing 40 billion to 150 billion total parameters, where it ranks first in the set of 67. The source says it ranks about 18th when all open-weight models are counted. It also describes Super VL as the top-scoring permissively licensed US open-weight model; that qualification matters because the source says some alternatives have different licensing terms.

Infrastructure changes the speed picture. The 630-token-per-second result came from one pre-release endpoint using NVFP4 weights, while the provider-wide figures combine six services. The gap between those results indicates that hardware, provider and serving setup affect what users may experience. Teams considering deployment need to test their intended endpoint, prompt lengths and workload rather than assume the fastest reported result will apply.

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

— Thorsten Meyer, ThorstenMeyerAI.com

Production Costs and Performance Vary

The published figures do not establish how Super VL will perform on a particular company’s tasks, or what its full production cost will be. The $0.25 estimate is tied to Artificial Analysis’ index evaluation. Prompt size, cached-input use, output length, provider pricing, hardware and operational overhead can change real costs.

The 630-token-per-second figure is from one pre-release endpoint; provider-wide reported speeds span 106 to 534 tokens per second at the 5th-to-95th percentile. It is not clear from the source whether the same speed will be available consistently after launch or across different hardware and serving configurations. The benchmark results also do not answer whether the model is reliable enough for any specific high-impact workflow.

Finally, the supplied material does not provide a full account of the model’s performance across all benchmark categories, independent production evaluations or the practical terms involved in self-hosting. Buyers should verify licensing and deployment requirements for their own use rather than treating the source’s broad license description as legal advice.

Test the Intended Deployment

The immediate next step for prospective users is to compare Super VL on the provider or NVIDIA hardware they expect to run, using representative prompts and the same quality checks they apply to alternatives. That testing should capture end-to-end latency, input and output token volumes, cache behavior and failure rates—not just peak generation speed or list prices.

Teams weighing the model against other options can compare the relevant license, measured task quality and total serving cost side by side. Further provider measurements and independent evaluations would help clarify how consistently the model’s published speed and benchmark improvements carry over to production. The source material does not identify a scheduled follow-up evaluation or another announced product milestone.

Key Questions

What is Nemotron 3.5 Super VL?

It is an NVIDIA open-weight model released on October 15, 2026. The source describes it as a 120-billion-parameter mixture-of-experts model with 12 billion active parameters, a one-million-token context window and NVFP4 weights.

Does the model make AI tasks cheaper?

Artificial Analysis estimates a cost of $0.25 per Intelligence Index task, compared with $1.64 for Nemotron 3 Super. That is a benchmark-specific estimate, not a promise about production bills; real costs depend on prompts, caching, outputs and provider or hardware choices.

Is it the fastest model available?

The source reports about 630 output tokens per second on one pre-release DeepInfra endpoint for 100,000-token prompts. Across six providers, the reported median for that prompt length is 307 tokens per second. These figures do not establish a universal speed ranking across every model and deployment.

Does it offer the best intelligence score for the price?

Not according to the comparisons in the source. MiMo-V2.6-Pro is listed at 46 points for $0.13 per task, while Super VL is listed at 26 points for $0.25. The models may differ in licensing, speed and deployment options, so benchmark value is only one comparison.

What should organizations check before using it?

Test the intended provider or self-hosted setup with representative workloads. Measure quality, latency, token use and total costs, and review the license and deployment requirements for the organization’s specific use.

Source: ThorstenMeyerAI.com