Nemotron 3.5 Super VL: What Its Price And Performance Show

🔍 Read the full analysis: Nemotron 3.5 Super VL: What Its Price And Performance Show 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, and Artificial Analysis reported a score of 26.5 on its Intelligence Index, alongside a listed cost of $0.25 per index task. The 120-billion-parameter open-weight model is positioned as a fast, commercially usable option, but the benchmark figures show trade-offs: its measured speed depends heavily on the provider, and its score trails several alternatives that cost the same or less.

The model has 120 billion total parameters and 12 billion active in a mixture-of-experts design, takes text and images as input, and has a reported one-million-token context window. NVIDIA released weights in NVFP4, a four-bit format. The source lists the license as OpenMDW-1.1 and says it permits commercial use. The listed API prices are $0.20 per million input tokens and $0.575 per million output tokens, with cached input at $0.055 per million.

Artificial Analysis ranks Super VL first among 67 open-weight models in its 40-billion-to-150-billion-parameter size class. That is not the same as leading all open models: the source places it around 18th across open-weight models overall. Its reported $0.25 cost per index task is lower than the $1.64 listed for its predecessor, Nemotron 3 Super, while the new model’s index score is twice as high, at 26.5 versus 13.

The speed results need qualification. A pre-release DeepInfra endpoint running the NVFP4 weights reportedly reached about 630 output tokens per second on prompts of 100,000 tokens. Across six providers, however, Artificial Analysis reports a median of 196 tokens per second on its standard prompt and 307 on 100,000-token prompts; provider results ranged from 106 to 534 tokens per second. The median time to the first token was 0.89 seconds, while the source reports about 11 seconds of reasoning before the first answer token and 13.6 seconds for a 500-token answer.

Where Speed and Cost Pay Off

For teams running repeated document, repository or agent workloads, the headline price per output token may not determine the bill. The supplied analysis estimates that about 89% of the $0.25 index-task cost comes from input, with cache writes alone accounting for 69%; output costs less than three cents in that task breakdown. Workloads that repeatedly send large inputs may benefit from the low cached-input rate, while tasks dominated by other usage patterns could see different costs.

The benchmark results also separate throughput from responsiveness. High throughput can help with batch jobs and background agents, but the reported 11-second reasoning period may be noticeable in interactive use. Teams should test the specific provider and workload they plan to use rather than treating the single-endpoint 630-token-per-second result as a guaranteed production speed.

Super VL’s appeal is not simply that it offers the most intelligence per dollar. The source lists open weights, commercial-use permission and the option to run the model on NVIDIA hardware as potential reasons to choose it. Those factors may matter to organizations weighing deployment control and licensing alongside benchmark scores. They do not, by themselves, establish that self-hosting is cheaper or faster for a particular team.

How It Compares With Alternatives

The source’s comparison suggests that Super VL is not the strongest listed choice on score per task cost. Xiaomi’s MiMo-V2.6-Pro is reported at 46 points for $0.13 per task, while MiMo-V2.6-Flash scores 38 for $0.06. DeepSeek V4.1 Flash is listed at 39 for $0.27, and GLM-5.3-Flash at 42 for $0.28. Inkling Small has the same reported score of 26 as Super VL, at $0.09 per task. These are figures reported in the source, not independent assessments here.

The source also compares the model with proprietary offerings: Claude Haiku 5.5 at maximum effort is listed at 43 points for $0.21 per task, and GPT-6 Luna at maximum effort at 38 points for $0.07. Comparisons can depend on model settings, task mix and pricing assumptions, so these numbers should not be treated as a complete evaluation of service quality or total deployment costs.

The licensing distinction affects one specific claim. Super VL is described as the top-scoring permissively licensed US open-weight model in the source’s comparison. Meta’s Muse Spark Open is reported at 39 points, but under a license the source says permits commercial use with restrictions. That makes the ranking relevant to buyers with particular licensing requirements, rather than a universal ranking of US models.

The source reports improvement over Nemotron 3 Super on selected measures: GDPval-AA, a benchmark for economically useful tasks, rose from 490 to 1,162 Elo, and AA-Briefcase rose from 876 to 999. AutomationBench-AA is reported at 14.8%, compared with 3.0% for Nemotron 3 Ultra. On AA-Omniscience, the reported hallucination rate fell from 87% to 36% compared with Nemotron 3 Super. These are benchmark-specific results, not proof of performance across all tasks.

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

— Thorsten Meyer, in the supplied source article

What the Benchmarks Do Not Settle

The supplied material does not establish how the model performs across every production workload, or whether the launch-period benchmark results will hold as providers update their systems. The 630-token-per-second result is tied to one pre-release endpoint; measured speeds across six providers varied substantially. The source advises testing the intended endpoint, but does not provide results from a broad set of customer deployments.

It is also unclear from the supplied figures how much self-hosting would cost after hardware, energy, staffing and serving capacity are included. The source identifies the option to run the weights on NVIDIA hardware but gives no full comparison of self-hosted and API costs. The benchmark descriptions also do not provide enough detail to resolve how differences in evaluation settings affect every model-to-model comparison.

The supplied account ends partway through its discussion of the model’s strengths and gaps. It gives selected benchmark results but does not provide a full assessment of image understanding, error types, robustness or safety. The reported reduction in hallucination on AA-Omniscience should be read as a result on that evaluation, not a guarantee that the model will hallucinate at the same rate in other settings.

Tests Before Production Deployment

The immediate next step for prospective users is to evaluate the model against their own tasks, using the provider, prompt lengths and caching patterns they expect to deploy. That testing can show whether the speed advantage applies to their setup and whether input-heavy pricing offsets the model’s other costs. Teams considering self-hosting will also need to calculate hardware and operating costs separately; the source provides no total-cost estimate.

Further evidence will depend on additional provider measurements and more extensive evaluations of real workloads. The supplied material does not announce a specific upcoming benchmark, release date or product milestone. Until that evidence is available, the figures support a measured conclusion: Super VL combines a competitive class-leading score, low reported task cost and strong speed on some endpoints, but it is not the highest-scoring or cheapest option in the comparisons presented.

Key Questions

What score did Nemotron 3.5 Super VL receive?

Artificial Analysis is reported to have given it 26.5 on the Intelligence Index. The source says it ranked first among 67 open-weight models in its 40-billion-to-150-billion-parameter size class, but around 18th among open-weight models overall.

How much does one Intelligence Index task cost?

The reported cost is $0.25 per task. That figure is an evaluation-based cost estimate, not a guarantee of what a specific production workload will cost.

Is the 630-token-per-second speed typical?

Not across the provider measurements supplied. The 630-token-per-second figure came from one pre-release DeepInfra endpoint; Artificial Analysis reported a six-provider median of 307 tokens per second on 100,000-token prompts, with results ranging from 106 to 534.

Does the license allow commercial use?

The source identifies the license as OpenMDW-1.1 and says commercial use is allowed. Organizations should review the license terms for their own intended use rather than relying on a short description.

Is Nemotron 3.5 Super VL the best value?

Not on the source’s score-and-cost comparisons alone. Several listed models score higher for similar or lower per-task costs, while Super VL’s case also depends on speed, licensing and the option to run its weights. Which option is best depends on workload and deployment requirements.

Source: ThorstenMeyerAI.com