How Fast Is Nemotron 3.5 Super VL? Under A Second

🔍 Read the full analysis: How Fast Is Nemotron 3.5 Super VL? Under A Second 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, with Artificial Analysis reporting an Intelligence Index score of 26, a $0.25 cost for one index task and output speeds ranging from a 196-token-per-second provider median to 630 tokens per second on one pre-release endpoint. The figures make the model a potentially fast, relatively low-cost option, but its measured performance depends on the provider, and its score is below several models listed at similar or lower task costs.

Artificial Analysis gives Nemotron 3.5 Super VL a score of 26.5, rounded to 26, twice the 13 recorded for NVIDIA’s Nemotron 3 Super. In the 40-billion-to-150-billion-total-parameter size class, the source says it ranks first among 67 open-weight models. That comparison is narrower than a ranking of all open-weight models: counting each model once across sizes, the source places it at about 18th. The model has 120 billion total parameters and 12 billion active in a mixture-of-experts design, supports a claimed one-million-token context window, and accepts text and images while producing text.

The headline speed figure came from a pre-release DeepInfra endpoint serving NVIDIA’s NVFP4, or four-bit, weights. It produced about 630 output tokens per second with prompts of 100,000 tokens. Artificial Analysis described this as the fastest result it had measured at that prompt length among models scoring 20 or more. The broader provider figures are lower: across six providers, the reported median was 196 output tokens per second on standard prompts and 307 with 100,000-token prompts. The reported range for standard output speed ran from 106 tokens per second at the fifth percentile to 534 at the 95th percentile.

Latency is another part of the release’s appeal, though it is distinct from the time needed to produce a full response. Artificial Analysis reports a 0.89-second median time to first token, compared with 1.76 seconds for Inkling Small and 1.23 for Nemotron 3 Ultra. The source says reasoning before the answer can take about 11 seconds, while a 500-token answer takes about 13.6 seconds end to end. Listed API prices are $0.20 per million input tokens and $0.575 per million output tokens; cached input is $0.055 per million tokens. Artificial Analysis puts the cost of an index task at $0.25, versus $1.64 for the predecessor, and reports that the new model generated about 46,000 output tokens per task compared with 114,000 for Nemotron 3 Super.

Speed Depends on the Endpoint

For teams running document-heavy agents or other workloads with large prompts, the reported low latency and comparatively modest benchmark-task cost could make Super VL useful. But the 630-token-per-second result is not a guarantee of what a production deployment will deliver: the six-provider median and wide reported spread point to meaningful differences between endpoints. Buyers should test the provider, hardware, prompt lengths and response sizes they plan to use rather than treating the fastest observation as a typical result.

The cost comparison also depends on workload. Artificial Analysis attributes about 89% of the $0.25 index-task cost to input, with cache writes alone accounting for 69% in its breakdown; output costs under three cents. That profile suggests the headline output-token rate may not determine an agent’s bill when it repeatedly reads large documents or repositories. Actual spending will depend on token use, caching and the application’s request pattern.

On the same index-task comparison, the source lists MiMo-V2.6-Pro at score 46 and $0.13 per task, MiMo-V2.6-Flash at 38 and $0.06, DeepSeek V4.1 Flash at 39 and $0.27, and GLM-5.3-Flash at 42 and $0.28. Inkling Small has the same score of 26 at a listed $0.09 per task. These are source-reported comparisons, not proof that the models will perform identically on a particular company’s work. They do show that Super VL’s case is not simply the highest intelligence score for the price; speed, deployment choices and licensing may matter more to some buyers.

A New Model Against Its Predecessor

The release follows NVIDIA’s Nemotron 3 Super. According to Artificial Analysis figures in the source material, the newer model doubles its predecessor’s Intelligence Index score from 13 to 26 and reduces the reported cost per index task from $1.64 to $0.25. It also uses fewer output tokens on that benchmark. Those results give buyers a direct comparison within NVIDIA’s model line, although benchmark performance does not establish how the model will handle every production task.

The source describes the weights as released in NVFP4 and the license as OpenMDW-1.1, with commercial use allowed. It characterizes the model as the top-scoring permissively licensed US open-weight model. That label has limits: the source says the ranking is specific to US models meeting its licensing definition, and compares it with Muse Spark Open, which scored 39 but uses a license it describes as commercially usable with restrictions. License terms and model provenance can affect deployment decisions independently of benchmark scores.

The reported performance profile is uneven. Artificial Analysis figures cited by the source show GDPval-AA, a benchmark of economically useful work, at 1,162 Elo, up from 490 for Nemotron 3 Super and above Nemotron 3 Ultra’s 1,013 and Inkling Small’s 1,124. The source also lists 999 Elo on AA-Briefcase, compared with 876 for Nemotron 3 Ultra. On AutomationBench-AA, it reports 14.8%, about five times Nemotron 3 Ultra’s 3.0%, while acknowledging that the absolute result remains low. These benchmark results provide context for the overall score; they do not establish accuracy across all business workflows.

Production Speed and Reliability

The reported speed is a set of benchmark observations, not a single performance guarantee. The fastest figure came from one pre-release endpoint, and the source does not provide enough detail here to establish whether production hardware, traffic, software settings or prompt composition would reproduce it. The six-provider spread also means an organization’s result may differ substantially from the median. Real-world throughput and latency remain workload- and provider-dependent.

The source material reports an AA-Omniscience hallucination rate falling from 87% for Nemotron 3 Super to 36% for Super VL. That is a benchmark-specific comparison; it does not mean errors occur at that rate in every application, nor does it show that the new model is reliable enough for unsupervised use. The model’s behavior on a particular company’s documents, image inputs and workflows remains to be tested.

Other details also warrant verification before deployment. The source lists a one-million-token context window, but does not provide a complete account here of performance across that full window. It does not specify the precise conditions behind every provider result, or provide a full comparison of operational costs beyond the index-task estimate and token prices. The licensing description says commercial use is allowed, but organizations should review the OpenMDW-1.1 terms directly for their intended use.

Test Before Production Deployment

The next practical step for prospective users is to benchmark Super VL on the endpoint and infrastructure they expect to use. Tests should reflect their typical prompt lengths, image inputs, cache behavior and answer sizes, and should measure time to first token, full-response time, throughput and total cost. That will help determine whether the model’s speed and pricing advantages hold for their own workload.

Teams should also evaluate task accuracy and hallucination risks against their existing systems, especially if outputs will be acted on without review. Comparing Super VL with alternatives on the same prompts and evaluation criteria can clarify whether its combination of speed, licensing and deployment options justifies choosing it over a higher-scoring or lower-cost model. The source material does not provide a later release milestone or a schedule for additional benchmark results, so further provider-specific and production measurements remain the next evidence to watch.

Key Questions

Does Nemotron 3.5 Super VL always run at 630 tokens per second?

No. The 630-token-per-second figure came from one pre-release DeepInfra endpoint using 100,000-token prompts. Artificial Analysis reported a six-provider median of 196 tokens per second on standard prompts and a wide provider range.

How long does it take to start answering?

Artificial Analysis reported a 0.89-second median time to first token. The source also says reasoning before an answer can take about 11 seconds and a 500-token response about 13.6 seconds end to end.

How much does an index task cost?

Artificial Analysis lists a cost of $0.25 per Intelligence Index task. The API prices in the source are $0.20 per million input tokens and $0.575 per million output tokens, with cached input at $0.055 per million. A particular application’s bill will depend on its token use and caching.

Is it the highest-scoring open-weight model?

The source says it ranks first among 67 open-weight models in its 40-billion-to-150-billion-parameter size class. Across open-weight models of all sizes, it places around 18th, so it is not described as the overall leader.

Can companies use the model commercially?

The source identifies the license as OpenMDW-1.1 and says commercial use is allowed. Organizations should review the license itself and confirm that its terms fit their intended deployment.

Source: ThorstenMeyerAI.com