Nemotron 3.5 Super VL: 25 Cents Per Task And Sub-Second Latency

🔍 Read the full analysis: Nemotron 3.5 Super VL: 25 Cents Per Task And Sub-Second Latency on ThorstenMeyerAI.com

Before you orderOffer from Amazon

Get movie nights 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 26.5 and an estimated $0.25 per benchmark task. The 120-billion-parameter model accepts text and images, is released with weights in NVFP4 format, and is offered under a license the source describes as allowing commercial use. Its results point to a fast, relatively low-cost option, but not a leading choice for raw benchmark intelligence at its price.

Artificial Analysis places the model first among 67 open-weight models in its 40-billion-to-150-billion-parameter size class, according to the supplied source. That ranking is limited to models in the specified class: across all open-weight models, the source puts it at about 18th. Its reported score is 26.5, compared with 13 for its predecessor, Nemotron 3 Super.

The model has 120 billion total parameters and 12 billion active parameters in a mixture-of-experts design, with a reported one-million-token context window. Listed API pricing is $0.20 per million input tokens and $0.575 per million output tokens; cached input is $0.055 per million tokens. Artificial Analysis estimates the cost of one Intelligence Index task at $0.25, compared with $1.64 for Nemotron 3 Super. Those figures describe the benchmark task and are not a universal cost for every application.

Speed depends on where the model is served. A pre-release DeepInfra endpoint running NVFP4 weights reportedly generated about 630 output tokens per second with 100,000-token prompts. Across six providers, Artificial Analysis reports a median of 307 tokens per second for that prompt length and 196 tokens per second on its standard prompt. The reported median time to first token is 0.89 seconds, but the source says reasoning adds about 11 seconds before the first answer token; a 500-token answer takes about 13.6 seconds end to end.

The Trade-Off Behind the Price

For teams choosing a model, the headline price is only one part of the calculation. The reported benchmark cost is low relative to Nemotron 3 Super, but comparisons in the source show that several alternatives combine higher Intelligence Index scores with similar or lower task costs. For example, Xiaomi’s MiMo-V2.6-Pro is listed at 46 points and $0.13 per task, while Inkling Small is listed at 26 points and $0.09. These are figures from the cited comparison, not guarantees of equivalent results on a team’s own workload.

Nemotron’s case is instead a combination of speed, NVIDIA hardware support, commercial-use licensing and the ability to run released weights on an organization’s own infrastructure. Those features may matter to teams with specific deployment, data-control or procurement requirements. The claimed throughput advantage also needs care: the 630-token figure came from one endpoint, while provider-level results vary substantially. Teams should test their intended provider and prompts before estimating production response times or bills.

Cost composition may matter as much as the per-token rates. The source’s breakdown attributes about 89% of the benchmark task cost to input, including 69% to cache writes, with output below three cents. That pattern could benefit from attention by operators processing long documents or repeatedly reading repositories: input and caching behavior may dominate their bill. Actual costs will depend on prompt lengths, cache use, output volume and provider pricing.

How the Benchmark Compares

The release follows Nemotron 3 Super, whose score was reported as 13 on the same Intelligence Index. The supplied source says the new model also reduced the predecessor’s cost per index task from $1.64 to $0.25 and used about 46,000 output tokens per task, compared with 114,000 for the earlier model. That is a benchmark-specific comparison, not a claim that every real-world workload will see the same reduction.

Several other results help show where the model sits. The source lists DeepSeek V4.1 Flash at 39 points and $0.27 per task, GLM-5.3-Flash at 42 and $0.28, and MiMo-V2.6-Flash at 38 and $0.06. It also reports Claude Haiku 5.5 at 43 points and $0.21, and GPT-6 Luna at 38 and $0.07, both at maximum effort. Benchmark scores and costs can depend on evaluation settings; the supplied material does not provide a full methodology for reconciling every model’s deployment conditions.

The licensing comparison is narrower than a general claim of being the best open model. The source describes Nemotron as a top-scoring option among permissively licensed US open-weight models, while saying Meta’s Muse Spark Open, released on the same day, scored 39 under a license with commercial-use restrictions. Whether a license meets a company’s needs depends on its terms and legal review.

“The country question is also less simple than the headline.”

— ThorstenMeyerAI.com

Provider Results and License Terms

The reported speed figures are not uniform: the highest figure comes from one pre-release endpoint, while the provider-wide median is lower and the reported range runs from 106 to 534 tokens per second at the 5th and 95th percentiles. The source does not establish what speeds users will receive under different hardware, traffic or service conditions. Production latency and throughput remain provider- and workload-dependent.

The supplied material does not include the complete benchmark methodology, the full terms of the OpenMDW-1.1 license, or independent production tests. It also does not provide detailed results for every evaluation category; for example, its discussion of hallucination reports a fall from 87% to 36% on AA-Omniscience but does not explain the metric’s full setup. The benchmark findings should not be treated as proof of performance on a particular company’s tasks.

What Teams Should Test Next

NVIDIA’s release makes the model and its weights available, while the source lists six providers. The next practical step for prospective users is to test the specific provider, quantization, prompts and workload they plan to use. That testing can establish response latency, output quality, image handling, actual token use and the effect of cached input on cost.

Organizations considering self-hosting will also need to check hardware capacity and review the license against their intended use. No later benchmark milestone or additional NVIDIA announcement is specified in the supplied material. Until more comparable testing is available, the reported figures offer a starting point rather than a production guarantee.

Key Questions

What is Nemotron 3.5 Super VL?

It is an NVIDIA-released 120-billion-parameter open-weight model with 12 billion active parameters. The supplied source says it accepts text and images, produces text, and has a one-million-token context window.

How much does it cost per task?

Artificial Analysis estimates $0.25 for one Intelligence Index task. That is a benchmark-specific estimate; costs for a real application depend on input and output tokens, caching, provider rates and workload.

Is the model’s speed always 630 tokens per second?

No. The approximately 630 output tokens per second result was reported for one pre-release DeepInfra endpoint using 100,000-token prompts. The reported median across providers for that prompt length was 307 tokens per second, with a wide provider range.

Does it outperform other models for the price?

Not on the comparisons provided. Several listed alternatives scored higher for similar or lower benchmark task costs. Nemotron’s potential advantages instead include its reported speed, licensing position and option to run its released weights on NVIDIA hardware.

What remains uncertain before production use?

Actual performance on a given workload, provider-level latency, full evaluation methodology, hardware requirements and how the license applies to a particular deployment are not established by the supplied material. Teams should run their own tests and review the license terms.

Source: ThorstenMeyerAI.com