Nemotron 3.5 Super VL: 25 Cents A Task, Sub-Second Latency, And A Score Of 26

🔍 Read the full analysis: Nemotron 3.5 Super VL: 25 Cents A Task, Sub-Second Latency, And A Score Of 26 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 assigning it an Intelligence Index score of 26. The model pairs a reported $0.25 cost per index task with fast performance on some endpoints, but that score does not make it the cheapest or highest-scoring choice among the models compared in the source report.

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. It accepts text and images and produces text. The released weights use NVIDIA’s NVFP4, a four-bit format, and the source lists an OpenMDW-1.1 license that permits commercial use. It reports six providers offering the model.

Artificial Analysis scores Nemotron 3.5 Super VL at 26, or 26.5 in the listed specification. The source says that places it first among 67 open-weight models in its 40-billion-to-150-billion-parameter size class. Across all open-weight models, however, it ranks about 18th, according to the report. Its score is twice the 13 given to its predecessor, Nemotron 3 Super.

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. The source estimates one Artificial Analysis Intelligence Index task costs $0.25, compared with $1.64 for Nemotron 3 Super. That task estimate reflects the evaluation’s token use, not a fixed price for every real-world request.

Where Speed and Cost Matter

The model’s strongest case is a mix of latency, deployment choice and licensing, rather than a claim to lead on general intelligence per dollar. The source reports a median time to first token of 0.89 seconds across providers, compared with 1.76 seconds for Inkling Small and 1.23 seconds for Nemotron 3 Ultra. That could suit automated workflows that benefit from quick starts and can run jobs in the background.

For interactive use, the reported first-token figure can be misleading without accounting for reasoning time. The source says the model takes about 11 seconds to produce its first answer token and around 13.6 seconds to complete a 500-token answer. Those timings may be acceptable in a pipeline but noticeable in a live chat. Teams should test their own prompts, provider and deployment setup before treating the headline latency as a user-facing response time.

Cost also depends on what a workload sends. In the source’s index-task breakdown, about 89% of the $0.25 bill comes from input, with cache writes alone accounting for 69%; output costs less than three cents. Applications that repeatedly submit long documents or codebases may therefore see costs driven more by input and caching than by the advertised output rate. The source’s prices are API figures; it does not provide a comparable total cost for operating the model on a company’s own hardware.

How the Model Compares

The reported 630 output tokens per second came from a pre-release DeepInfra endpoint using NVFP4 weights and prompts of 100,000 tokens. Artificial Analysis reportedly described it as the fastest model it had measured at that prompt length among models scoring at least 20. That result is not the same as typical provider performance: the source gives a six-provider median of 196 tokens per second on standard prompts and 307 with 100,000-token prompts, with a range from 106 to 534 at the 5th and 95th percentiles. Provider selection can materially change observed speed.

The report’s value comparisons include MiMo-V2.6-Pro at 46 points for $0.13 per task, MiMo-V2.6-Flash at 38 for $0.06, DeepSeek V4.1 Flash at 39 for $0.27, and GLM-5.3-Flash at 42 for $0.28. Inkling Small is listed at 26 for $0.09. These comparisons use the source’s Artificial Analysis scores and task-cost estimates; actual API bills can differ with prompt length, caching, output size and provider.

The license distinction affects the comparison too. The source says Nemotron 3.5 Super VL is the highest-scoring permissively licensed US open-weight model, while Meta’s Muse Spark Open scored 39 but uses a license with commercial-use restrictions. That makes the “top” claim dependent on the category being compared—not a ranking across all models or licenses.

“Super VL is a very good engine and an average brain.”

— Thorsten Meyer, ThorstenMeyerAI.com

What the Benchmarks Leave Open

The figures describe benchmark results and selected provider measurements, not a guarantee of performance on every task. The source says the model improved on several measures, including knowledge work and SaaS automation, but does not provide enough detail here to establish how those results transfer to particular business workloads. The report gives an AutomationBench-AA result of 14.8%, compared with 3.0% for Nemotron 3 Ultra; the higher result is a benchmark comparison, not evidence that most real-world tasks will be automated successfully.

The source also reports a reduction in hallucination rate on AA-Omniscience from 87% for Nemotron 3 Super to 36% for Super VL. That is a benchmark-specific measure, not proof that the model will avoid errors in deployment. The supplied material cuts off before giving further detail about the model’s weaknesses, so its failure patterns, image-task performance and reliability on less common inputs remain unclear here.

Other unknowns include the cost and performance of self-hosting, the exact hardware needed for production, and whether the reported API prices and provider availability will remain stable. The score appears as both 26 and 26.5 in the source, and readers should check the underlying evaluation page for the precise displayed value and methodology.

What Teams Should Test

Organizations considering the model should run their own prompts against the provider or hardware they intend to use, measuring full response time, token consumption, accuracy and failure rates. Long-context performance should be tested with the actual documents or repositories in scope, including how caching changes total input cost. For interactive products, teams should measure time to a useful answer—not only time to the first generated token.

Legal and infrastructure reviews will also shape adoption. Buyers should verify that the license terms fit their intended use, confirm the available deployment options and calculate total costs that include hardware, hosting and maintenance. The source reports commercial use is allowed under OpenMDW-1.1, but does not supply a detailed legal analysis of the license.

Further comparison will depend on updated provider measurements and independent evaluations across specific tasks. Until then, the reported results support a measured conclusion: Nemotron 3.5 Super VL may appeal to teams prioritizing speed, commercial-use terms or NVIDIA-based deployment, while its score and cost figures alone do not establish it as the best-value model for every workload.

Key Questions

What is Nemotron 3.5 Super VL’s Artificial Analysis score?

The source lists an Intelligence Index score of 26, also shown as 26.5 in its specification table. It says this is first among 67 open-weight models in the 40-billion-to-150-billion-parameter class, but about 18th across all open-weight models.

Does every task cost $0.25?

No. $0.25 is the source’s estimated cost for one Intelligence Index task. Real costs depend on input and output token counts, caching, provider pricing and the workload being run.

How fast is the model?

The source reports a 630-token-per-second result from one pre-release endpoint. Across six providers, it gives median output speeds of 196 tokens per second on standard prompts and 307 on 100,000-token prompts. It also reports a median time to first token of 0.89 seconds, while reasoning adds time before an answer appears.

Is it the best-value open model?

Not by the comparisons in the source. Several listed open-weight models score higher for similar or lower estimated task costs, and Inkling Small is listed with the same score of 26 for $0.09. Nemotron’s case instead rests on factors such as speed, license and deployment options.

Can companies use the model commercially?

The source identifies its 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