Nemotron 3.5 Super VL: Understanding Its Under-a-Second Latency

🔍 Read the full analysis: Nemotron 3.5 Super VL: Understanding Its Under-a-Second Latency 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, with Artificial Analysis reporting a median 0.89-second time to first token and an Intelligence Index score of 26.5. The model’s fast initial response, commercial-use license and reported $0.25 cost per benchmark task may interest teams deploying AI on NVIDIA hardware, but its speed varies by provider and the index score does not make it a top-value choice for every workload.

Artificial Analysis places the model at 26.5 on its Intelligence Index, roughly double the 13 scored by its predecessor, Nemotron 3 Super. It ranks first among 67 open-weight models in its total-parameter size class, defined in the source as 40 billion to 150 billion parameters. Across open-weight models of all sizes, the source puts it at about 18th. The model has 120 billion total parameters, 12 billion active in a mixture-of-experts design, and accepts text and images as input while producing text.

Speed figures need to be read with their test conditions. A pre-release DeepInfra endpoint serving NVIDIA’s NVFP4 weights generated about 630 output tokens per second on prompts of 100,000 tokens. Artificial Analysis described that as its fastest measurement at that prompt length among models scoring at least 20. Across six providers, however, the model page reports a median of 307 tokens per second with 100,000-token prompts, with results ranging from 106 to 534. Median output speed on the standard prompt was 196 tokens per second.

The reported 0.89-second median is the time to the first token under the benchmark’s measurement; it is not the time to receive a complete answer. The source says the model reasons before responding, with the first answer token taking about 11 seconds and a 500-token response about 13.6 seconds end to end. API prices are listed as $0.20 per million input tokens and $0.575 per million output tokens, with cached input at $0.055 per million. Artificial Analysis estimates a benchmark-task cost of $0.25, compared with $1.64 for Nemotron 3 Super.

Latency Depends on the Endpoint

For production teams, a fast first response can improve the feel of an interactive tool and reduce delays before an agent starts returning useful content. But time to first token is not completion time: the source’s estimate of roughly 11 seconds to the first answer token, followed by more generation time, means the subsecond figure alone cannot predict how quickly a user receives a finished response.

Provider variation also matters for operational planning. The gap between 106 and 534 output tokens per second across providers is large enough that a published peak should not be treated as a guaranteed deployment speed. Teams considering the model should benchmark the precise provider, hardware, prompt length and response type they expect to use. For self-hosted deployments, the ability to run released weights on NVIDIA hardware may be more relevant than a hosted endpoint’s result.

The price comparison also depends on workload. Artificial Analysis says about 89% of the $0.25 task cost is input, with cache writes accounting for 69% and output costing less than three cents. That profile suggests repeated document or repository input could shape an agent’s bill more than its generated answer. The model is cheaper than its predecessor on the cited benchmark, but other models in the source score higher for similar or lower task costs.

Products Worth Considering

Apple MacBook Air 15-inch Laptop with M5 Chip, 16GB Memory, 512GB SSD, Midnight

AmazonView Latest Price

Jackery Explorer 2000 v2 Portable Power Station with Two 200W Solar Panels

AmazonView Latest Price

Bose QuietComfort Headphones (2nd Gen), Wireless Over-Ear Noise-Cancelling Headphones, Black

AmazonView Latest Price

As an affiliate, we earn on qualifying purchases.

A Faster Successor, With Trade-Offs

Nemotron 3.5 Super VL is presented as a successor to Nemotron 3 Super, with a higher Artificial Analysis Intelligence Index score and a lower estimated cost per index task. Its listed context window is one million tokens, and the released weights use the NVFP4 format. The source says the model is available through six providers and is licensed under OpenMDW-1.1, which it describes as allowing commercial use.

The source also reports improvements on selected evaluations, not just the overall index. GDPval-AA, which tests economically useful work, rose to 1,162 Elo from 490 for Nemotron 3 Super. AA-Briefcase scored 999 Elo, compared with 876 for Nemotron 3 Ultra. AutomationBench-AA reached 14.8%, versus 3.0% for Nemotron 3 Ultra, although the source notes that the absolute result remains low. On AA-Omniscience, the reported hallucination rate declined from 87% to 36% compared with the predecessor.

Those results sit alongside alternatives that the source says score higher on the index. For example, it lists MiMo-V2.6-Pro at 46 for $0.13 per task and Inkling Small at 26 for $0.09. The comparison is not a universal ranking for every use: licensing terms, deployment options, modality, reliability and task-specific performance can change which model fits a team’s needs. The source describes Super VL’s advantage as including its speed, licensing and self-hosting option, rather than raw benchmark value alone.

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

— Thorsten Meyer, ThorstenMeyerAI.com

What the Benchmarks Do Not Show

The figures supplied are benchmark results and provider measurements, not a guarantee of performance in a particular product. It is not clear from the source how the 0.89-second median was measured across prompts and providers, or whether it applies to the same conditions as the 630-token-per-second result. The source gives provider throughput percentiles but does not provide a full breakdown of latency and pricing for each provider.

It is also unclear how the model performs across a broad range of real-world visual tasks. Although it is a vision-language model and accepts images, the supplied results focus largely on text-oriented index tasks and selected evaluations. The source does not give enough detail to establish how its hallucination rate, automation score or answer quality will transfer to a particular use case.

Comparisons depend on benchmark versions, inference settings and licensing requirements. The reported cost per task is an Artificial Analysis estimate for its index evaluation, not a universal cost per user request. The source’s claim that the model leads among permissively licensed US open-weight models is also specific to its stated comparison category; it does not mean the model leads all open models or all commercial options.

Test the Intended Deployment

The next practical step for teams considering Nemotron 3.5 Super VL is to test it on their own prompts and infrastructure. That should include time to first token, total completion time, input and cache costs, output quality and performance on image-based tasks if those are part of the workload. A provider-specific trial can show whether the published speed and pricing fit the service’s needs better than the benchmark averages do.

Further comparisons will depend on additional provider measurements, independent evaluations and details about deployment conditions. Until those are available, the current figures support a narrower conclusion: Super VL has a strong reported latency and cost improvement over its predecessor, but neither its fastest endpoint result nor its benchmark score alone settles whether it is the right production model.

Key Questions

What does under-a-second latency mean here?

Artificial Analysis reports a 0.89-second median time to first token. That measures the wait for the initial token, not for a complete answer; the source estimates about 11 seconds to the first answer token and 13.6 seconds for a 500-token response.

Is 630 tokens per second the typical speed?

No. The 630-token-per-second result came from one pre-release DeepInfra endpoint with 100,000-token prompts. Across six providers, the reported median for prompts of that length was 307 tokens per second, with a range from 106 to 534.

How much does Nemotron 3.5 Super VL cost?

The listed API rates are $0.20 per million input tokens and $0.575 per million output tokens. Cached input is listed at $0.055 per million tokens. Artificial Analysis estimates a $0.25 cost for one of its Intelligence Index tasks; that is a benchmark-specific estimate, not a fixed price for every request.

Does it outperform other open models?

It ranks first in the source’s 40-billion-to-150-billion total-parameter open-weight size class, but about 18th among open-weight models overall. The source lists several models with higher Intelligence Index scores, including MiMo-V2.6-Pro and DeepSeek V4.1 Flash.

Can companies use the model commercially?

The source says the model uses the OpenMDW-1.1 license and permits commercial use. Organizations should review the license itself and confirm that its terms match their deployment and legal requirements.

Source: ThorstenMeyerAI.com