What The 26 Score Says About Nemotron 3.5 Super VL

🔍 Read the full analysis: What The 26 Score Says About Nemotron 3.5 Super VL on ThorstenMeyerAI.com

Before you orderOffer from Amazon

Get hardware and tech essentials 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, 2026, with an Artificial Analysis Intelligence Index score of 26—double the reported 13 for its predecessor, Nemotron 3 Super. The score puts it at the top of the source’s 40-billion-to-150-billion-parameter open-weight size class, but other reported measures show a more qualified picture: the model is positioned around speed, cost and commercial-use licensing, while its overall benchmark score trails several alternatives.

Artificial Analysis lists Super VL as a 120-billion-parameter mixture-of-experts model, with 12 billion parameters active, a context window of up to one million tokens and weights released in NVFP4, a four-bit format. Its listed API prices are $0.20 per million input tokens and $0.575 per million output tokens; cached input is listed at $0.055 per million tokens. The source reports a cost of about $0.25 for one Intelligence Index task.

The cost comparison with Nemotron 3 Super is substantial in that benchmark: the earlier model is reported at $1.64 per index task, with a score of 13. The article attributes the lower cost partly to reduced output volume—about 46,000 output tokens per task, versus 114,000 for the predecessor—and to lower cached-input pricing. It says roughly 89% of the new model’s task cost comes from input, including cache writes, while output accounts for less than three cents.

Speed depends on where the model is served. A pre-release DeepInfra endpoint using NVFP4 weights produced about 630 output tokens per second with 100,000-token prompts, according to the source’s account of Artificial Analysis measurements. Across six providers, the model page’s reported median was 196 tokens per second on a standard prompt and 307 with 100,000-token prompts, with a 106-to-534 range. Median time to the first token was 0.89 seconds; the source says a reasoning answer’s first response token took about 11 seconds, and a 500-token answer took about 13.6 seconds end to end.

Speed, Cost and Benchmark Trade-offs

The score is useful as a point of comparison, not a complete verdict on whether a model is suitable for a product. Super VL’s reported $0.25 per index task is below the reported class median, and its lower cost than Nemotron 3 Super arrives alongside a higher score. But the source’s comparisons show that low cost and a 26 score are not unique: Inkling Small is reported at the same score for about $0.09 per task, while Xiaomi’s MiMo-V2.6-Flash scores 38 at $0.06. Other listed alternatives score higher at similar or lower task costs.

That makes Super VL’s potential appeal more specific than “best intelligence for the money.” Teams may place weight on its reported speed, NVIDIA hardware deployment, model provenance and license, especially if they need control over hosting or have legal requirements for permissive terms. Those considerations can matter for production systems, but they need to be weighed against actual task performance, endpoint speed and input-token usage. The source’s cost breakdown suggests that systems repeatedly processing large documents or repositories should pay particular attention to input and cache costs.

Latency also changes how the model may feel to users. A quick first token in a median provider measurement does not mean the full reasoning response is immediate: the source reports an approximately 11-second wait before the answer begins in its described reasoning scenario. That may be acceptable for background jobs or multi-step agents, but it can be noticeable in interactive chat. The appropriate comparison depends on the application’s tolerance for delay, not just tokens per second.

How the Benchmark Compares

The Intelligence Index score is a composite benchmark measure, and the source places Super VL at number one among 67 open-weight models in its stated size class, defined as 40 billion to 150 billion total parameters. That class-specific ranking is narrower than a ranking across all open-weight models: the article estimates Super VL at about 18th when models are counted once across the broader group. The two rankings describe different comparison sets and should not be conflated.

The source reports gains over Nemotron 3 Super on several task measures. GDPval-AA, which evaluates economically useful work, is listed at 1,162 Elo, up from 490 for the predecessor, and above the cited figures for Nemotron 3 Ultra and Inkling Small. AA-Briefcase is reported at 999 Elo, compared with 876 for Ultra. AutomationBench-AA is listed at 14.8%, versus 3.0% for Ultra; the source notes that the newer model’s result remains low in absolute terms.

Another reported change is on AA-Omniscience: the hallucination rate fell from 87% to 36% compared with Nemotron 3 Super. That is a benchmark-specific result, not a guarantee that the model will avoid errors in deployment. The article also says Super VL leads among permissively licensed U.S. open-weight models, while noting that other models may score higher under licenses with additional restrictions.

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

— Thorsten Meyer, ThorstenMeyerAI.com

Provider Results and Real-World Fit

The available figures do not establish how Super VL will perform across all production workloads. The cited 630-token-per-second result came from one pre-release DeepInfra endpoint, while the six-provider median and range differ substantially. Provider hardware, serving configuration, prompt length and workload can affect results; the source does not provide enough information here to predict an individual team’s latency or costs.

The score and task prices also do not answer whether the model will perform well on a particular company’s documents, tools or evaluation set. The reported hallucination rate is tied to AA-Omniscience, and the source does not give deployment-level error rates. Nor does the material specify the complete terms of the OpenMDW-1.1 license beyond saying that commercial use is allowed. Organizations should review the license itself and test their intended use rather than treating the summary as legal advice.

Finally, the source’s claim that Super VL leads its specified permissive U.S. model category depends on the category and licensing definition used. A comparison with models under more restrictive licenses is not a like-for-like legal comparison. The benchmark rankings, costs and license distinctions should be read with those limits in mind.

What Teams Should Test Next

The practical next step for prospective users is to test the model through the provider and hardware configuration they expect to use. That should include representative prompt lengths, first-token delay, full response time, throughput under concurrent requests and the share of input that can be cached. Since the source reports that most index-task cost comes from input, teams should calculate costs using their own document and conversation patterns rather than relying only on per-token headline prices.

Teams should also compare Super VL against alternatives on the same tasks and scoring rules, including measures of accuracy, tool use, hallucinations and completion time. The reported benchmark results provide a starting point, but they do not replace evaluations on a company’s own workload. License review is another step for organizations considering commercial deployment or self-hosting.

The source material does not specify a subsequent NVIDIA release date or a planned benchmark update. Further provider measurements and independent evaluations could clarify how consistent the speed and cost figures are outside the cited tests.

Key Questions

What does a score of 26 mean for Nemotron 3.5 Super VL?

It is the model’s reported Artificial Analysis Intelligence Index score. The source says that score is double Nemotron 3 Super’s 13 and ranks Super VL first within the stated 40-billion-to-150-billion-parameter open-weight class. It is not a score for every possible task or a guarantee of performance on a particular workload.

Is Nemotron 3.5 Super VL the fastest model?

The source reports about 630 output tokens per second on one pre-release DeepInfra endpoint with 100,000-token prompts. Across six providers, the reported median for that prompt length was 307 tokens per second, with results ranging from 106 to 534. The 630 figure should not be treated as a speed users will receive from every provider.

How much does one benchmark task cost?

Artificial Analysis is reported as estimating a cost of about $0.25 per Intelligence Index task. The source says roughly 89% of that figure comes from input, including cache writes. Actual costs can vary with provider, prompt size, caching and output volume.

Can businesses use the model commercially?

The source identifies its license as OpenMDW-1.1 and says commercial use is allowed. Businesses should check the full license terms for their specific deployment and legal requirements.

Does the score show that Super VL is the best-value model?

No. The source lists several models with higher scores at similar or lower benchmark task costs, and Inkling Small at the same score for a lower reported cost. Super VL may appeal for other reasons, including speed, deployment options and licensing, but value depends on the task and operating requirements.

Source: ThorstenMeyerAI.com