The Numbers Behind Nemotron 3.5 Super VL’s AI Performance

🔍 Read the full analysis: The Numbers Behind Nemotron 3.5 Super VL’s AI Performance on ThorstenMeyerAI.com

Buying for a business?Offer from Amazon

Get business pricing on office and shipping supplies

• Business-only prices and quantity discounts
• Tax-exempt purchasing
• Multiple users, one account, clear invoices

Create a free Business account →Free to register

As an affiliate, we earn on qualifying purchases.Hide

NVIDIA released Nemotron 3.5 Super VL on October 15, 2026, and Artificial Analysis rated it 26.5 on its Intelligence Index, double the 13 recorded for Nemotron 3 Super. The figures describe a fast, relatively inexpensive open-weight model, but they do not make it the strongest value by benchmark score: several models listed in the source score higher for similar or lower estimated task costs.

Artificial Analysis places Super VL at No. 1 among 67 open-weight models in its 40-billion-to-150-billion-parameter size class. That ranking is class-specific: across all open-weight models counted once, the source places it at about 18th. The model has 120 billion total parameters and 12 billion active in a mixture-of-experts design, a one-million-token context window, and accepts text and images while producing text. NVIDIA released its weights in NVFP4, a four-bit format. The source lists the OpenMDW-1.1 license and says commercial use is allowed.

Speed depends substantially on where the model is hosted. A pre-release DeepInfra endpoint serving the NVFP4 weights produced about 630 output tokens per second with 100,000-token prompts, according to the source’s account of Artificial Analysis measurements. The model page’s median across six providers was lower: 196 tokens per second on a standard prompt and 307 with 100,000-token prompts. The reported provider range for the latter measure was 106 to 534 tokens per second. The figures are not interchangeable: the 630 result came from one endpoint, rather than the cross-provider median.

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 a complete Intelligence Index task at $0.25. In its launch write-up, the organization put the predecessor’s cost at $1.64 per task, making the new model about 85% less costly on that measure while scoring twice as high. The source says Super VL used about 46,000 output tokens per index task, versus 114,000 for Nemotron 3 Super, and that input accounted for about 89% of its estimated task cost.

Speed Comes With Trade-Offs

The results matter to teams choosing a model for long-document processing, image input or automated workflows. A one-million-token context window and relatively low listed prices may suit workloads that repeatedly supply large documents or code repositories. But the source’s cost breakdown indicates that input and cached-input usage can weigh more heavily on an index task’s bill than output generation. Buyers should compare costs using their own prompt lengths, caching patterns and traffic rather than relying on the per-token output price alone.

For interactive products, the speed figures need a second qualification: throughput is not the same as the time a user waits for a complete answer. The reported median time to first token is 0.89 seconds, but the source says the model takes about 11 seconds to produce its first answer token because it reasons before responding. A 500-token answer took about 13.6 seconds end to end in the reported measurements. That profile may be acceptable for background agents or batch pipelines, but it could feel slow in a live chat.

The benchmark comparison also limits what “good value” means here. The source 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 are Artificial Analysis comparisons, not guarantees of performance on a particular application. Super VL’s case may instead rest on speed, the stated license, NVIDIA hardware compatibility and the option to run the weights on a team’s own infrastructure.

How the Benchmarks Compare

The headline ranking depends on the comparison group. Artificial Analysis’ 40B-to-150B parameter class puts Super VL first among 67 open-weight models, while the source estimates its position at about 18th when open-weight models of all sizes are counted. Those descriptions can both be true; neither means it leads every model on every task.

The supplied comparisons include DeepSeek V4.1 Flash at 39 for $0.27 per task and GLM-5.3-Flash at 42 for $0.28. It also lists Claude Haiku 5.5 at 43 for $0.21 and GPT-6 Luna at 38 for $0.07, both at maximum effort. These figures are benchmark and cost estimates cited in the source, and the models differ in licensing and deployment options. Super VL’s particular licensing distinction is that it is described as the leading US open-weight option with a permissive license; the source says Meta’s Muse Spark Open scored 39 but has commercial-use restrictions under its Muse Model License.

The source also reports gains in selected task evaluations. On GDPval-AA, described as measuring economically useful tasks, Super VL scored 1,162 Elo, compared with 490 for Nemotron 3 Super. Its AA-Briefcase score was 999 Elo, up from 876 for Nemotron 3 Ultra. On AutomationBench-AA, the reported score was 14.8%, compared with 3.0% for Nemotron 3 Ultra. Those measures offer a more specific view than the overall index, but the source notes that the automation result remains low in absolute terms.

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

— Thorsten Meyer, ThorstenMeyerAI.com

Provider Results and Open Questions

The supplied material does not include a complete set of Super VL evaluation results. It reports improvements on GDPval-AA, AA-Briefcase and AutomationBench-AA, and says the hallucination rate on AA-Omniscience fell from 87% to 36% compared with Nemotron 3 Super. However, it provides no further methodology or task-level detail for that hallucination measure, and the information ends before describing other strengths, failures or evaluation results. These figures should not be treated as a full account of reliability.

It is also unclear how closely the reported costs and throughput will match a particular deployment. The listed task cost is an Artificial Analysis estimate, while the speed measurements differ by provider and endpoint. The material does not specify the compute requirements or operating costs for running the model on an organization’s own NVIDIA hardware, nor does it provide independent evidence about performance on customers’ real workloads. The license is described as permitting commercial use, but organizations should review the actual license terms for their intended use.

Test Against Your Workload

The next practical step for prospective users is to test their intended provider and model configuration, using representative prompts, images, context lengths and caching behavior. Teams should measure both throughput and full response time, then compare the quality and cost against alternatives on the tasks they actually need to complete. The source identifies six providers, but does not report a later update or a schedule for additional evaluations.

For organizations considering self-hosting, the relevant follow-up is whether their available NVIDIA hardware can serve the NVFP4 weights at the required scale and cost. No deployment capacity figures are provided here. More provider-specific results, complete evaluation details and real-world operating data would help clarify how far the launch benchmarks carry over to production use.

Key Questions

What did Nemotron 3.5 Super VL score?

Artificial Analysis gave it 26.5 on the Intelligence Index, double the 13 reported for Nemotron 3 Super. The source says it ranked first among 67 open-weight models in its 40B-to-150B parameter size class.

Is Super VL the best-value model in the comparisons?

Not by Intelligence Index score per estimated task cost in the comparisons supplied. The source lists models such as MiMo-V2.6-Pro with a higher score and lower estimated task cost. Super VL may appeal for other reasons, including its stated license, speed and self-hosting option.

How fast is it?

The source reports about 630 output tokens per second on one pre-release DeepInfra endpoint with 100,000-token prompts. Across six providers, the cited median at that prompt length was 307 tokens per second, with a reported range of 106 to 534.

What does an Intelligence Index task cost?

Artificial Analysis estimated a task at $0.25. The source says input accounted for about 89% of that estimate, so actual costs can depend on prompt size and caching as well as output volume.

What remains unconfirmed about production use?

The supplied material does not establish performance or operating costs on a reader’s specific hardware and workload. It also does not provide a full set of task-level evaluations or detailed methodology for the reported hallucination-rate change.

Source: ThorstenMeyerAI.com