Three Metrics For Nemotron 3.5 Super VL: Price, Speed, Score

🔍 Read the full analysis: Three Metrics For Nemotron 3.5 Super VL: Price, Speed, Score on ThorstenMeyerAI.com

Age 18–24?Offer from Amazon

Prime made for students and young adults

• Fast, free delivery for dorm and study essentials
• Prime Video and Amazon Music included
• Member-only deals

Try Prime for Young Adults →Free trial for eligible 18–24 year olds

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 on its Intelligence Index, alongside pricing and speed results. The model is a 120-billion-parameter open-weight system with 12 billion active parameters; the figures point to a fast, relatively low-cost option, but not a leading choice for intelligence per dollar.

Artificial Analysis lists the model at $0.20 per million input tokens and $0.575 per million output tokens, with cached input priced at $0.055 per million tokens. It estimates $0.25 per Intelligence Index task. That is substantially below the $1.64 per task reported for its predecessor, Nemotron 3 Super, which scored 13. NVIDIA’s model uses NVFP4 weights, and its OpenMDW-1.1 license permits commercial use, according to the supplied source material.

Speed depends on where the model runs. A pre-release DeepInfra endpoint produced about 630 output tokens per second on prompts of 100,000 tokens, a result Artificial Analysis described as the fastest it had measured at that prompt length among models scoring at least 20. Across six providers, however, the reported median was 307 tokens per second on 100,000-token prompts and 196 tokens per second on the standard prompt. Provider results ranged from 106 to 534 tokens per second at the 5th and 95th percentiles.

The benchmark report also gives a median 0.89-second time to first token. It says the model takes about 11 seconds to produce its first answer token when reasoning, and about 13.6 seconds to complete a 500-token response. The source reports a major improvement in hallucination rate on AA-Omniscience, from 87% for Nemotron 3 Super to 36% for Super VL; these are benchmark results, not a guarantee about performance in every deployment.

Cost and Speed Have Trade-Offs

The results matter to teams deciding whether to use a hosted model or run open weights on their own infrastructure. Nemotron 3.5 Super VL combines commercially permitted licensing, low listed token prices and relatively fast generation. Its 120-billion-parameter total size and 12-billion active-parameter mixture-of-experts design may also make it relevant to operators with suitable NVIDIA hardware, though the provided material does not give a full self-hosting cost comparison.

Its task price is not the same as best value for benchmark intelligence. The source compares it with MiMo-V2.6-Pro, which scored 46 at $0.13 per task, and Inkling Small, which scored 26 at $0.09. Other listed models, including DeepSeek V4.1 Flash and GLM-5.3-Flash, also scored higher for a similar task price. Those comparisons rely on Artificial Analysis’s reported benchmark and pricing figures; real costs can vary with provider, workload and usage pattern.

Input use is a key part of the bill. Artificial Analysis’s task breakdown puts about 89% of the $0.25 cost on input, with cache writes accounting for 69% and output costing less than three cents. Teams whose systems repeatedly submit long documents or code repositories should pay attention to caching and prompt size, rather than judging cost from the output-token rate alone.

How the Benchmark Rankings Fit

The Intelligence Index score of 26.5, rounded to 26 in the source’s headline, is reported as double Nemotron 3 Super’s 13. Artificial Analysis ranks Super VL first among 67 open-weight models in its stated size class, defined as models with 40 billion to 150 billion total parameters. That class-specific result should not be read as a ranking across all open-weight models: the source places Super VL at about 18th overall when counting each model once.

The source also identifies strengths beyond the overall score. On GDPval-AA, a benchmark for economically useful work, Super VL recorded 1,162 Elo, compared with 490 for Nemotron 3 Super, 1,013 for Nemotron 3 Ultra and 1,124 for Inkling Small. Its AA-Briefcase result was 999 Elo, up from 876 for Nemotron 3 Ultra. On AutomationBench-AA it scored 14.8%, compared with 3.0% for Ultra, while the source cautions that the newer score remains low in absolute terms.

The license comparison needs care. The source says Super VL leads among permissively licensed US open-weight models. It cites Muse Spark Open, which scored 39, but says its Muse Model License permits commercial use with restrictions. A team’s legal requirements can change which comparison is relevant; license terms should be checked directly before deployment.

“Super VL is not the best value for raw intelligence, even among open models.”

— Thorsten Meyer, ThorstenMeyerAI.com

Provider Results and Real Costs

The reported speed figures do not establish how fast every user will experience the model. The 630-token-per-second result came from one pre-release endpoint, while provider-wide figures varied considerably. The source does not specify how those results change across hardware configurations, production traffic or different workload mixes, so buyers should test the provider and prompt lengths they expect to use.

The supplied material does not detail the full methodology behind every benchmark, confidence ranges, or how the reported task costs translate to a particular application. It also gives no complete comparison of self-hosting expenses, such as hardware acquisition, operation and staffing. Benchmark improvements in knowledge work, automation and hallucination are specific to the named tests; they do not establish equivalent results in a customer’s own system.

Several comparisons depend on current model prices and licenses, which can change. The source provides no update on later provider availability or changes to the model and benchmark listings. The difference between a permissive license and one with commercial restrictions also requires review of the actual terms, not just a ranking label.

Test the Intended Deployment

Teams considering the model should compare it on their own tasks, including long-context inputs and any required image processing, and measure end-to-end response time rather than relying only on peak generation speed. They should also model input, cache and output charges against expected usage, then verify the hosting provider’s available precision and hardware options.

For organizations that need commercial permissions or control over where model weights run, the next step is to review the OpenMDW-1.1 terms and assess self-hosting requirements. For organizations choosing primarily on benchmark score per dollar, Artificial Analysis’s listed alternatives provide a reason to test competing models side by side. Updated provider measurements and independent results on practical workloads would help clarify how the launch figures translate to production.

Key Questions

What score did Nemotron 3.5 Super VL receive?

Artificial Analysis reported an Intelligence Index score of 26 (listed as 26.5 in the source’s specification table). The source says it ranks first among 67 open-weight models in the 40-billion-to-150-billion-parameter size class, not across all open-weight models.

How much does it cost per benchmark task?

Artificial Analysis estimates $0.25 per Intelligence Index task. Its 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 tokens. Actual spending depends on usage and provider.

How fast is the model?

A pre-release DeepInfra endpoint reached about 630 output tokens per second on 100,000-token prompts. Across six providers, the reported median at that prompt length was 307 tokens per second, with results ranging from 106 to 534 tokens per second at the 5th and 95th percentiles.

Is it the best value among open models?

Not by the score-and-cost comparisons in the supplied source. It lists models such as MiMo-V2.6-Pro with a higher score and lower estimated task cost, and Inkling Small with the same score for less. Super VL’s case also includes speed, license and the possibility of running its weights on suitable NVIDIA hardware.

What remains to be tested before production use?

Users should verify performance on their own workload and chosen provider, including latency, long-input costs and output quality. The published benchmarks and one endpoint’s peak speed do not establish results for every deployment.

Source: ThorstenMeyerAI.com