Nemotron 3.5 Super VL: From 25-Cent Tasks To A Score Of 26

🔍 Read the full analysis: Nemotron 3.5 Super VL: From 25-Cent Tasks To A Score Of 26 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 gave it an Intelligence Index score of 26. The model offers a 120-billion-parameter mixture-of-experts design, a reported $0.25 cost per index task and weights released in NVFP4, giving developers a fast, lower-cost option they can deploy on NVIDIA hardware.

Artificial Analysis lists 120 billion total parameters, with 12 billion active, and a context window of up to 1 million tokens. It accepts text and images and produces text. Its OpenMDW-1.1 license allows commercial use, according to the source material. API prices are listed as $0.20 per million input tokens and $0.575 per million output tokens; cached input is $0.055 per million tokens. The model page covers six providers.

The source reports a sharp improvement over Nemotron 3 Super, which scored 13 and cost $1.64 per Intelligence Index task. Super VL’s reported task cost is about 85% lower. Artificial Analysis’s launch write-up puts output at roughly 46,000 tokens per index task, compared with 114,000 for its predecessor. The source’s cost breakdown estimates that input accounts for about 89% of the $0.25 task bill, including 69% from cache writes.

Speed results vary by measurement. A pre-release DeepInfra endpoint serving NVFP4 weights reached about 630 output tokens per second on prompts of 100,000 tokens. Across six providers, the reported median was 307 tokens per second at that prompt length, with results ranging from 106 to 534. Median time to first token was 0.89 seconds, but the source says reasoning takes about 11 seconds before the first answer token; a 500-token response took about 13.6 seconds end to end.

Speed Comes With a Value Trade-Off

For teams running long-context workflows, throughput, self-hosting and commercial-use terms may matter as much as benchmark scores. The source reports that Super VL cut its predecessor’s task cost while doubling its Intelligence Index score, and says the model can run on NVIDIA hardware. Its lower cached-input price could also benefit systems that repeatedly process large documents or code repositories, though actual bills depend on the workload’s input and cache use.

On the source’s comparison, however, a score of 26 is not a leading result for the price. It lists MiMo-V2.6-Pro at 46 for $0.13 per task, MiMo-V2.6-Flash at 38 for $0.06, and Inkling Small at 26 for $0.09. These comparisons make the choice workload-specific: Super VL’s reported speed and license may suit some operators, while other models offer higher scores or lower task costs. The supplied figures do not establish that every model was measured under identical deployment conditions.

How the Score Is Ranked

The score of 26 comes from Artificial Analysis’s Intelligence Index, a benchmark measure cited in the source. The source says Super VL ranks first among 67 open-weight models in its total-parameter size class, defined as 40 billion to 150 billion parameters. Across open-weight models overall, it places the model around 18th. The size-class result is a narrower comparison than an overall ranking.

The source also describes Super VL as the leading permissively licensed US open-weight model. It contrasts it with Muse Spark Open, which it says scored 39 but uses a license with commercial-use restrictions. The ranking and licensing comparison are attributed here to the supplied source material; the available details do not include the underlying benchmark methodology or license text.

Deployment Results Still Vary

The 630-token-per-second result came from one pre-release provider endpoint, while the six-provider median was substantially lower. The source does not specify the final production configuration for that endpoint or how each provider’s hardware and serving setup affected performance. Operators should treat the reported range as evidence that provider choice changes observed speed, rather than assume the top figure will be typical.

The supplied source material also stops partway through its evaluation breakdown, after reporting knowledge-work, SaaS automation and hallucination results. It says the hallucination rate on AA-Omniscience fell from 87% for Nemotron 3 Super to 36% for Super VL, but does not provide the full benchmark details or results for other areas. The available information does not establish how the model performs across a particular organization’s tasks, nor its costs under that organization’s prompt and caching patterns.

Test the Intended Workload

The practical next step for teams considering deployment is to measure the provider and configuration they plan to use. That means checking latency, sustained throughput and task costs with representative prompts, especially when workloads include long documents or repeated inputs. The source’s cross-provider spread makes endpoint-specific measurement relevant to any estimate based on the launch figures.

Further evaluation details would help clarify how Super VL’s reported gains extend beyond the metrics included in the source. Until those results and workload-specific tests are available, the evidence supports a model with a 26-point index score, reported low task cost and variable speed; it does not settle whether it is the best fit for a particular production system.

Key Questions

When did NVIDIA release Nemotron 3.5 Super VL?

NVIDIA released it on October 15, 2026, according to the supplied source.

What score did it receive?

Artificial Analysis gave it an Intelligence Index score of 26. The source also reports a score of 26.5 in its specification list.

How much does one index task cost?

The source reports a cost of $0.25 per Intelligence Index task. Actual costs for other workloads will depend on token use and caching.

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.

Can businesses use the model commercially?

The supplied source says its OpenMDW-1.1 license allows commercial use. Teams should review the license terms for their intended deployment.

Source: ThorstenMeyerAI.com