Nemotron 3.5 Super VL Shows A Score Of 26 And Fast Response Times

🔍 Read the full analysis: Nemotron 3.5 Super VL Shows A Score Of 26 And Fast Response Times on ThorstenMeyerAI.com

Before you orderOffer from Amazon

Get audio and creator gear 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 an Intelligence Index score of 26, twice the 13 scored by its predecessor, Nemotron 3 Super. The results put the new model at the top of the 40B–150B open-weight size class in that benchmark, while reported speed varied substantially by provider.

The model has 120 billion total parameters, with 12 billion active in its mixture-of-experts architecture, and accepts text and images as input. Its stated context window is 1 million tokens. NVIDIA released NVFP4 weights, and the source reports an OpenMDW-1.1 license that allows commercial use. Six providers offer the model.

Artificial Analysis lists API prices of $0.20 per million input tokens and $0.575 per million output tokens, with cached input at $0.055 per million. It estimates the cost of one Intelligence Index task at $0.25. The source says that is about 85% below Nemotron 3 Super’s $1.64 per task, while the newer model scored twice as high. Those cost figures reflect the benchmark’s evaluation workload, not every user’s likely bill.

Speed results depend on the measurement. A pre-release DeepInfra endpoint serving NVFP4 weights produced about 630 output tokens per second with 100,000-token prompts. Artificial Analysis’ provider-wide page instead reports a median of 307 tokens per second at that prompt length, with a range from 106 to 534 at the fifth and 95th percentiles. Its median first-token time is 0.89 seconds; the source estimates about 11 seconds to the first answer token when reasoning is included and 13.6 seconds for a 500-token response.

Speed Comes With Trade-Offs

For teams running repeated inference, the model’s price, latency and license may matter as much as its benchmark score. The reported $0.25 per index task is low relative to Nemotron 3 Super, and the model can be run using released weights. Those features could suit background agents or high-volume workflows where response time and deployment control are priorities.

The benchmark does not establish that Super VL is the best-value model for general reasoning. The source lists MiMo-V2.6-Pro at 46 points for $0.13 per task, DeepSeek V4.1 Flash at 39 for $0.27, and Inkling Small at 26 for $0.09. These comparisons are specific to the cited Artificial Analysis index and cost estimates. They do not capture every deployment constraint, but they show why a score of 26 alone cannot settle which model is the best fit.

Reported task results also suggest an uneven profile. The source cites gains in knowledge-work and SaaS automation evaluations, alongside a reduction in hallucination rate relative to Nemotron 3 Super. These measures can help teams decide what to test, but benchmark outcomes do not guarantee equivalent performance on a particular product or dataset.

How the Benchmarks Compare

Artificial Analysis places Super VL first among 67 open-weight models in its 40B–150B total-parameter class, according to the source. That ranking is narrower than a claim that it leads all open-weight models: the article says its overall position is about 18th when each model is counted once. The comparison depends on the benchmark’s model grouping and scoring methodology.

The source reports that Nemotron 3 Super scored 13 on the same index. It also gives Super VL a score of 26.5 in its specification table, while referring to the headline result as 26. This article uses 26 as the rounded score and preserves the source’s more precise figure where relevant. For operational measures, the source reports GDPval-AA at 1,162 Elo, up from 490 for the predecessor; AutomationBench-AA at 14.8%, compared with 3.0% for Nemotron 3 Ultra; and an AA-Omniscience hallucination rate of 36%, down from 87% for Nemotron 3 Super.

The licensing comparison also has limits. The source describes Super VL as the leading US open-weight model with a permissive commercial-use license. It contrasts Meta’s Muse Spark Open, which it says scored 39 but uses a license with commercial restrictions. That distinction concerns licensing terms, not a direct measure of model quality.

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

— Thorsten Meyer, ThorstenMeyerAI.com

Provider Results Still Vary

The 630-token-per-second figure came from one pre-release endpoint. Artificial Analysis’ six-provider figures show a lower median and a wide spread, so the speed a customer receives will depend on the provider and setup. The source does not identify the precise conditions needed to reproduce the endpoint result across providers.

The source’s cost estimates are tied to benchmark tasks and listed token prices. Actual bills will depend on prompt length, caching, output volume and provider pricing. It says input accounts for about 89% of the cited task cost, but does not establish that this split applies to every workload. Independent replication of the reported benchmark results and operational performance is not described.

Some evaluation details are also absent from the supplied material, including uncertainty ranges for scores and the full methodology behind comparisons. The reported reduction in hallucinations is a benchmark result; how often errors occur in a specific use case remains to be measured.

Test Before Deployment

The next practical step for prospective users is to measure the model on the provider and hardware they intend to use. Tests should reflect their own prompt lengths, image inputs, caching patterns and response requirements, since the source reports substantial provider differences in throughput and a gap between first-token latency and time to a reasoned answer.

Teams weighing the model against alternatives can compare both benchmark results and the tasks that matter to them, including knowledge work, automation and hallucination handling. The supplied source gives no date for further evaluations or a future NVIDIA release, so the next confirmed milestone is not specified.

Key Questions

What score did Nemotron 3.5 Super VL receive?

Artificial Analysis reported a rounded Intelligence Index score of 26. The source’s specification table gives a more precise figure of 26.5.

How fast is Nemotron 3.5 Super VL?

A pre-release DeepInfra endpoint recorded about 630 output tokens per second with 100,000-token prompts. Artificial Analysis reports a six-provider median of 307 tokens per second at that prompt length, so results vary by endpoint.

How much does it cost to use through an API?

The source lists prices of $0.20 per million input tokens, $0.575 per million output tokens and $0.055 per million cached input tokens. It estimates $0.25 for one Artificial Analysis index task; actual costs depend on usage and provider.

Is the model commercially licensed?

The source identifies its license as OpenMDW-1.1 and says commercial use is allowed. Users should check the license terms for their intended deployment.

Does a score of 26 make it the best-value open model?

No single index score establishes best value for every workload. The source lists several open models with higher scores at similar or lower benchmark-task costs, while Super VL’s case includes speed, licensing and the option to run released weights.

Source: ThorstenMeyerAI.com