Is Nemotron 3.5 Super VL Fast Enough For Sub-Second Tasks?

🔍 Read the full analysis: Is Nemotron 3.5 Super VL Fast Enough For Sub-Second Tasks? on ThorstenMeyerAI.com

Before you orderOffer from Amazon

Get smart everyday buys 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, with Artificial Analysis publishing benchmark results the same day that show a model with 0.89-second median time to first token but a much longer wait for its first answer token when reasoning is included. Its measured throughput ranges widely across providers, so whether it suits sub-second tasks depends on what “fast” means and where it runs.

Artificial Analysis gives Super VL an Intelligence Index score of 26.5 and places it first among 67 open-weight models in its 40-billion-to-150-billion-parameter size class. Across open-weight models overall, the source says it ranks about 18th. The model has 120 billion total parameters and 12 billion active in a mixture-of-experts design, accepts text and images, and has a stated context window of one million tokens.

Speed results depend on the measurement. A pre-release DeepInfra endpoint serving NVFP4 weights produced about 630 output tokens per second with 100,000-token prompts. Artificial Analysis describes that as its fastest measurement at that prompt length for a model scoring at least 20. Across six providers, however, the model page reports median output speeds of 196 tokens per second on its standard prompt and 307 with 100,000-token prompts; observed output speed spans 106 to 534 tokens per second.

The median time to first token is 0.89 seconds, but that is not the same as time to the first answer after reasoning. The supplied analysis puts the first answer token at about 11 seconds and a 500-token response at about 13.6 seconds end to end. Pricing is listed at $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.

When Fast Is Fast Enough

The distinction between first-token latency and a usable answer matters for teams considering sub-second interactions. A response that begins within a second may still leave a user waiting roughly 11 seconds before reasoning produces its first answer token. That could be acceptable for a background agent or a pipeline that can work asynchronously, but it may feel slow in a live chat or any workflow requiring an immediate decision.

Provider choice also affects whether benchmark speed translates to an application. The reported 630-token-per-second result came from one pre-release endpoint, while measurements across providers differ substantially. Teams should test the actual deployment route with their own prompt lengths and output requirements rather than treating the highest figure as a general service guarantee. The supplied material does not provide a separate end-to-end latency distribution for every provider.

Cost and license may outweigh raw benchmark rank for some operators. Artificial Analysis puts Super VL below its size-class median price, and the source says commercial use is allowed under OpenMDW-1.1. It also reports that input makes up most of the estimated index-task cost. The model may suit organizations prioritizing deployment control, NVIDIA hardware, or a permissive commercial-use license, even if another model offers a higher benchmark score per dollar.

How the Benchmark Compares

Super VL’s score is double the 13 reported for its predecessor, Nemotron 3 Super. The source also reports that the new model costs about 85% less per index task than its predecessor: $0.25 compared with $1.64. It attributes the difference in part to lower output volume on the index task, about 46,000 tokens compared with 114,000, and lower cached-input pricing.

The index score does not capture every strength. The supplied analysis reports a GDPval-AA score of 1,162 Elo, up from 490 for Nemotron 3 Super, and an AutomationBench-AA result of 14.8%, compared with 3.0% for Nemotron 3 Ultra. It also reports that the hallucination rate on AA-Omniscience fell from 87% to 36%. These are benchmark results cited by the source, not guarantees of performance on a particular workload.

Several comparisons complicate the value claim. The source lists open models MiMo-V2.6-Pro at 46 for $0.13 per task, MiMo-V2.6-Flash at 38 for $0.06, and DeepSeek V4.1 Flash at 39 for $0.27. Inkling Small is listed at 26 for $0.09. Those figures use the article’s reported benchmark and task-cost comparisons; practical results may differ by provider, configuration, and workload.

What the Speed Figures Leave Open

The published figures do not establish how Super VL will perform on every deployment or task. The 630-token-per-second result is tied to one pre-release endpoint; provider-level conditions, hardware, prompt size, and serving configuration can change results. The source gives a broad speed range but does not specify a guaranteed rate for production users.

It is also unclear whether the reported 11-second reasoning wait applies consistently across prompts and settings. No workload-specific results are provided for tasks that must finish inside one second, nor is there a full comparison of end-to-end latency across the six providers. The benchmark scores and reported task costs are useful reference points, but they do not settle which model will be more reliable or economical for a particular application.

Test the Intended Deployment

For teams evaluating Super VL, the next practical step is to benchmark the provider and hardware they intend to use, with representative prompt lengths and response sizes. Measure time to first token separately from time to first useful answer and total completion time; also track input, cached-input, and output usage to estimate costs for the actual workload.

The supplied material does not identify a future release date or a scheduled follow-up benchmark. Further provider measurements and independent tests could clarify how stable the speed figures are in production. Until then, the available evidence supports a qualified conclusion: Super VL can start generating quickly, but the reported reasoning delay and provider spread mean it should not be treated as a general sub-second answer model.

Key Questions

Is Nemotron 3.5 Super VL fast enough for sub-second tasks?

Its median time to first token is 0.89 seconds, but the reported time to its first answer token after reasoning is about 11 seconds. It may meet a sub-second target for initial output, but the figures do not show that it reliably completes useful answers within one second.

How fast is it across providers?

Artificial Analysis reports median output speeds of 196 tokens per second on the standard prompt and 307 tokens per second with 100,000-token prompts. Its reported range is 106 to 534 tokens per second, while the separate 630-token-per-second result came from one pre-release DeepInfra endpoint.

What does Nemotron 3.5 Super VL cost?

The listed prices are $0.20 per million input tokens, $0.575 per million output tokens, and $0.055 per million cached input tokens. Artificial Analysis estimates $0.25 for one Intelligence Index task; actual costs depend on usage and task composition.

Is it the highest-scoring open-weight model?

No. The source says it ranks first among 67 models in its 40-billion-to-150-billion-parameter size class, but about 18th among open-weight models overall. It also lists several models with higher index scores.

Can businesses use it commercially?

The supplied material says Super VL uses the OpenMDW-1.1 license and allows commercial use. Organizations should review the license terms themselves before deployment; this article does not provide legal advice.

Source: ThorstenMeyerAI.com