Nemotron 3.5 Super VL Posts Under-a-Second Latency And A 26 Score

🔍 Read the full analysis: Nemotron 3.5 Super VL Posts Under-a-Second Latency And A 26 Score 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 reported an Intelligence Index score of 26.5, alongside median time to first token of 0.89 seconds. The results put the model first among 67 open-weight models in its stated size class, according to the source material, but do not make it the highest-scoring open model overall: its benchmark position and practical speed depend on the comparison group and provider.

The model has 120 billion total parameters, with 12 billion active in its mixture-of-experts design, and accepts text and images as input. The source lists a one-million-token context window and weights released in NVFP4, a 4-bit format. It says the OpenMDW-1.1 license allows commercial use. The source does not provide the license text, so teams should check its terms directly before deployment.

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. Its estimate for one Intelligence Index task is $0.25. That figure is tied to the benchmark’s task mix, not a guaranteed cost for a particular production workload; actual bills depend on input volume, caching, output length and provider.

Throughput figures vary. A pre-release DeepInfra endpoint reportedly generated about 630 output tokens per second with 100,000-token prompts. Across six providers, however, the model page gives a median of 307 tokens per second at that prompt length, with a range from 106 to 534. The source reports 0.89 seconds median to first token; it also says a 500-token answer took about 13.6 seconds end to end, including roughly 11 seconds of reasoning before the answer began.

Speed Comes With Provider Variation

For developers running document-heavy pipelines or background agents, the reported short time to first token and low benchmark-task cost may make Super VL worth testing. Its large context window could also suit tasks involving long documents or code repositories, though the source provides no independent production measurements showing how the model performs on a specific team’s workload.

The results are less decisive for interactive chat. An approximately 11-second wait before answer text may feel slow even when the first-token metric is under a second, because the model is reported to reason before responding. And the 630-token-per-second result came from one endpoint, not the six-provider median. Teams evaluating the model should test the exact provider and settings they expect to use, rather than treating the peak figure as a universal service level.

The score also needs its comparison framed carefully. The source says 26.5 ranks first in the 40-billion-to-150-billion-parameter open-weight class, but places it about 18th when all open-weight models are counted once. It lists several models with higher scores at similar or lower estimated cost per task. Super VL’s case therefore rests on a combination of speed, licensing and self-hosting options, rather than a clear lead in intelligence per dollar.

How the 26.5 Score Compares

The source compares Super VL with its predecessor, Nemotron 3 Super, which scored 13 on the Artificial Analysis Intelligence Index. It reports that the new model costs about $0.25 per index task compared with $1.64 for the predecessor, and produces about 46,000 output tokens per task versus 114,000. Those are benchmark-specific estimates; they do not establish equivalent savings for every application.

Among alternatives cited in the source, MiMo-V2.6-Pro scored 46 at an estimated $0.13 per task, while MiMo-V2.6-Flash scored 38 at $0.06. DeepSeek V4.1 Flash scored 39 at $0.27, and GLM-5.3-Flash scored 42 at $0.28. Inkling Small scored 26 at $0.09. These comparisons are reported figures from the supplied material, not a separate verification of the models, prices or benchmark conditions.

The source also describes performance gains in selected tests: GDPval-AA rose to 1,162 Elo from 490 for Nemotron 3 Super, and AutomationBench-AA reached 14.8%, compared with 3.0% for Nemotron 3 Ultra. It reports the hallucination rate on AA-Omniscience falling from 87% to 36%. These are benchmark results, not guarantees of accuracy or success on real-world deployments.

“The 630 figure is real, but it came from one endpoint.”

— Thorsten Meyer, ThorstenMeyerAI.com

Benchmark Conditions Still Matter

The figures in the supplied report do not establish how the model will perform across production workloads, or whether the same latency and cost results will hold as providers change hardware, quantization or service settings. The spread between the reported 106 and 534 tokens per second across providers makes that variability material. The report also does not describe the complete benchmark methodology or provide independent verification of all listed comparisons.

It is also unclear from the source how the one-million-token context window performs at its maximum length, what trade-offs NVFP4 introduces for accuracy, or how the quoted task cost translates to a team’s own prompts and caching patterns. The reported reduction in hallucinations is a benchmark result, not evidence that the model is free of errors. The supplied material does not include a direct statement from NVIDIA or the full license terms.

Test the Intended Deployment

Teams considering the model should verify the license and deployment requirements, then benchmark it on their own prompts using the provider or NVIDIA hardware they plan to run. Useful checks include time to first useful answer, total response time, input and cache costs, accuracy on the team’s tasks, and behavior with long image or text inputs. Results should be compared with alternatives under the same conditions.

The next useful evidence will be broader provider measurements and independent evaluations of real-world performance, including the effects of the NVFP4 weights. The source material does not specify a date for further releases, benchmark updates or NVIDIA guidance, so no next milestone is confirmed.

Key Questions

What happened with Nemotron 3.5 Super VL?

NVIDIA released the model on October 15, 2026. Artificial Analysis reported an Intelligence Index score of 26.5, along with latency, throughput and estimated cost figures.

What does the score of 26 mean?

The source reports a score of 26.5 and says the model ranked first among 67 open-weight models in its 40-billion-to-150-billion-parameter size class. It also places the model about 18th among open-weight models overall, so the class ranking is not an overall lead.

Is the model under one-second latency?

The reported median time to first token is 0.89 seconds. That is not the same as receiving a complete answer: the source says reasoning can take about 11 seconds before answer text begins and a 500-token response took about 13.6 seconds end to end.

How much does it cost to use?

The listed API prices are $0.20 per million input tokens and $0.575 per million output tokens, with cached input at $0.055 per million. Artificial Analysis estimates $0.25 per index task; real costs depend on workload and provider.

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

Not according to the comparisons in the supplied report. It cites several open models with higher scores at similar or lower estimated task costs. Super VL may still suit teams that prioritize its reported speed, license terms or ability to run weights on their own NVIDIA hardware.

Source: ThorstenMeyerAI.com