A Score Of 26 And Sub-Second Latency For Nemotron 3.5 Super VL

🔍 Read the full analysis: A Score Of 26 And Sub-Second Latency For Nemotron 3.5 Super VL 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 published evaluation data the same day showing an Intelligence Index score of 26, a median 0.89-second time to first token and an estimated $0.25 per index task. The figures make the model a potentially fast, relatively inexpensive option for teams using open weights, but the score is below several alternatives available for similar or lower per-task costs.

Artificial Analysis lists the model at 120 billion total parameters, with 12 billion active in its mixture-of-experts design, and a one-million-token context window. It accepts text and images and returns text. NVIDIA released weights in NVFP4 format; the stated license, OpenMDW-1.1, allows commercial use. The source report says six providers offer the model.

The listed API prices are $0.20 per million input tokens and $0.575 per million output tokens. Cached input is priced at $0.055 per million tokens. Artificial Analysis estimates one of its Intelligence Index tasks costs $0.25. The report compares that with $1.64 for Nemotron 3 Super, its predecessor, which scored 13. It attributes the lower task cost in part to fewer generated tokens: about 46,000 per index task versus 114,000 for the earlier model.

Speed depends on the measurement and provider. A pre-release DeepInfra endpoint produced about 630 output tokens per second with 100,000-token prompts, according to the report. Across six providers, the model page’s median was 196 tokens per second on the standard prompt and 307 with 100,000-token prompts; the reported range was 106 to 534 tokens per second. Median time to first token was 0.89 seconds. The report says reasoning delays the first answer token to about 11 seconds, while a 500-token answer took about 13.6 seconds end to end.

Fast Starts, Uneven Value

The results distinguish initial latency from time to a reasoned answer. A median 0.89 seconds before the first token may suit interactive systems, but the reported 11 seconds before the first answer token could be noticeable in a chat product. For background agents and batch workflows, that wait may matter less than total throughput or the ability to process long prompts.

The cost figure also needs to be read as a task estimate, not a universal bill. The report says roughly 89% of the $0.25 task cost comes from input, with cache writes accounting for 69%; output costs less than three cents. Teams repeatedly sending large documents or repositories should measure their own input and caching patterns rather than rely on output-token prices alone.

On the report’s comparisons, the model’s score of 26 is not a clear raw-intelligence bargain. Xiaomi’s MiMo-V2.6-Pro scored 46 at $0.13 per task, while MiMo-V2.6-Flash scored 38 at $0.06. DeepSeek V4.1 Flash scored 39 at $0.27, and Inkling Small matched 26 at $0.09. These comparisons are Artificial Analysis figures reported by ThorstenMeyerAI.com, not guarantees of performance on every workload.

How the Score Is Framed

The score comes from the Artificial Analysis Intelligence Index, a benchmark measure rather than a direct prediction of performance in a particular company’s application. The report says Nemotron 3.5 Super VL leads a size category of 67 open-weight models with 40 billion to 150 billion total parameters. Across open-weight models of all sizes, it places around 18th, according to the same account. The “top” description therefore applies to a defined size class, not every open-weight model.

The release also improves on NVIDIA’s Nemotron 3 Super in the report’s comparisons: the index score doubles from 13 to 26, and the estimated task cost falls from $1.64 to $0.25. Some category results show larger changes. GDPval-AA, covering economically useful work, rises to 1,162 Elo from 490; AutomationBench-AA is 14.8%, compared with 3.0% for Nemotron 3 Ultra. On AA-Omniscience, the reported hallucination rate falls from 87% to 36%. Those benchmark gains do not establish performance across all real-world tasks.

The report describes the model as the highest-scoring permissively licensed US open-weight model, while noting that licensing conditions affect the comparison. Meta’s Muse Spark Open scored 39, but the report says its license permits commercial use with restrictions. Model scores, prices and licensing terms are separate considerations, and each can affect a deployment decision.

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

— Thorsten Meyer, ThorstenMeyerAI.com

Provider Results and Real-World Fit

The reported throughput figures do not describe a single consistent service experience: the 630-token-per-second result came from one pre-release DeepInfra endpoint, while the six-provider median was 196 tokens per second on the standard prompt. The source does not provide enough detail here to explain all differences among provider hardware, configuration or measurement conditions. Actual performance on a selected endpoint may differ.

The evaluation figures also leave open how the model will perform on a specific workload, including image-heavy tasks. Although the model accepts images, the supplied material does not give a separate visual benchmark result. Nor does it establish whether the reported hallucination reduction will carry over to individual production applications. Index scores and benchmark results are evidence about tested evaluations, not proof of outcomes in every deployment.

Pricing and licensing should be checked against the provider and use case a team intends to adopt. The source gives listed rates and describes the license as allowing commercial use, but it does not spell out all terms or deployment costs. It also does not establish that every provider offers the same speed, price or access to the released weights.

Test Before Production Use

The next practical step for prospective users is to run their own evaluations on the provider, prompts and workload they plan to use. Tests should track time to the first useful answer, total completion time, input and cached-token usage, output volume, task accuracy and image performance where relevant. This can show whether the model’s latency and cost profile outweigh its lower index score for a particular application.

Teams considering self-hosting will also need to assess their NVIDIA hardware, the NVFP4 weights and the operational costs of running the model. The supplied report does not provide a full self-hosting cost comparison or a schedule for future releases. For now, the confirmed development is the release and the initial Artificial Analysis evaluation; broader production results remain dependent on independent testing and provider-specific measurements.

Key Questions

What score did Nemotron 3.5 Super VL receive?

Artificial Analysis reported an Intelligence Index score of 26, or 26.5 in the detailed specification cited by ThorstenMeyerAI.com. The same report lists its predecessor, Nemotron 3 Super, at 13.

Does “sub-second latency” mean it answers in under a second?

Not necessarily. The reported 0.89 seconds is the median time to the first token. The report says reasoning delays the first answer token to about 11 seconds, and a 500-token answer took about 13.6 seconds end to end.

How fast is the model across providers?

The report gives a six-provider median of 196 output tokens per second on the standard prompt and 307 with 100,000-token prompts. A separate pre-release DeepInfra endpoint recorded about 630 tokens per second with the long prompt; that result is not the multi-provider median.

What does one Artificial Analysis task cost?

The reported estimate is $0.25 per Intelligence Index task. This is a benchmark-task cost, not a fixed price for every user request. The report says input makes up most of that task’s estimated cost.

Is the model suitable for commercial use?

The source identifies its license as OpenMDW-1.1 and says commercial use is allowed. Businesses should review the license terms and check the conditions attached to the specific provider or deployment they plan to use.

Source: ThorstenMeyerAI.com