How The Numbers Define Nemotron 3.5 Super VL

🔍 Read the full analysis: How The Numbers Define Nemotron 3.5 Super VL on ThorstenMeyerAI.com

Buying for a business?Offer from Amazon

Get business pricing on tech for your team

• 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, with Artificial Analysis reporting an Intelligence Index score of 26.5, a $0.25 cost per index task and high output speeds on a pre-release endpoint. The model leads the 40B-to-150B total-parameter class of open-weight models in that index, but comparisons in the supplied figures show that several models with similar or lower task costs score higher.

Super VL has 120 billion total parameters and 12 billion active parameters in a mixture-of-experts design, and accepts text and images while producing text. Its listed context window is one million tokens. NVIDIA released weights in NVFP4, a four-bit format, and the source lists six providers. The stated license, OpenMDW-1.1, allows commercial use. These specifications describe the release; performance and price can vary by provider and deployment.

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 full Intelligence Index task at $0.25, compared with $1.64 for Nemotron 3 Super. The source attributes much of the lower task cost to fewer generated tokens: about 46,000 per index task versus 114,000 for the predecessor. It says input accounts for roughly 89% of the new model’s task cost, including cache writes, while output costs less than three cents.

The release figures also show substantial variation in speed. A pre-release DeepInfra endpoint delivered about 630 output tokens per second on prompts of 100,000 tokens. Across six providers, however, Artificial Analysis reports a median of 307 tokens per second at that prompt length, with a range from 106 at the fifth percentile to 534 at the 95th percentile. Median output speed on its standard prompt was 196 tokens per second. Median time to first token was 0.89 seconds, but the source says reasoning before the answer can take about 11 seconds and a 500-token response about 13.6 seconds end to end.

The Trade-Off Behind a Score of 26

The benchmark results point to a model whose appeal is operational rather than purely score-based. Super VL is faster on some measured endpoints than the comparison models cited in the source, costs less per index task than its predecessor and can be run using released weights. For teams that need control over deployment, commercial-use rights or low latency before generation begins, those factors may matter as much as an aggregate intelligence score.

But the supplied comparisons weaken any claim that Super VL offers the most intelligence for the money. MiMo-V2.6-Pro scores 46 for a reported $0.13 per task, MiMo-V2.6-Flash scores 38 for $0.06, and Inkling Small scores 26 for $0.09. DeepSeek V4.1 Flash scores 39 for $0.27, while GLM-5.3-Flash scores 42 for $0.28. These are figures reported in the source and depend on Artificial Analysis’s task and pricing assumptions. They suggest buyers should compare their own workload, provider and license needs rather than treating one index score as a deployment decision.

The license comparison also needs care. The source calls Super VL the top-scoring permissively licensed US open-weight model, while noting that Meta’s Muse Spark Open scored 39 under a license with commercial-use restrictions. That distinction may matter to organizations whose legal review requires permissive terms. It does not make Super VL the highest-scoring open-weight model overall.

From Nemotron 3 Super to Super VL

The supplied launch analysis compares Super VL with Nemotron 3 Super, its predecessor, which scored 13 on the Intelligence Index. The new model’s 26.5 score is approximately double that figure. Its reported task cost also fell from $1.64 to $0.25. Those changes make the release a substantial step within NVIDIA’s Nemotron line, while leaving its absolute standing dependent on which other models and deployment costs are included in a comparison.

The source says the new model improved on several specific evaluations. Its GDPval-AA score, for tasks described as economically useful, rose to 1,162 Elo from 490, above the cited scores for Nemotron 3 Ultra and Inkling Small. AA-Briefcase is listed at 999 Elo, compared with 876 for Ultra. On AutomationBench-AA, Super VL scored 14.8%, versus 3.0% for Ultra; the source characterizes that as roughly five times higher while acknowledging the absolute score remains low. On AA-Omniscience, the reported hallucination rate fell from 87% to 36% versus the predecessor.

Artificial Analysis places Super VL first among 67 open-weight models in its 40B-to-150B total-parameter class. The source says it ranks about 18th when all open-weight models are counted once. These are different comparison groups, so the class-leading result should not be read as an overall open-weight ranking.

Provider Results and Real Workloads

The reported results do not establish how Super VL will perform on a particular company’s tasks. The source gives a wide spread in provider speeds and a headline figure drawn from one pre-release endpoint; actual latency and cost depend on the selected provider, prompt size, caching and output length. The listed $0.25 task cost is tied to Artificial Analysis’s index evaluation, not a universal price for every production request.

The source does not provide detailed results for every evaluation, nor does it specify how the reported hallucination rate translates to different applications. It also does not establish that every team can run the released NVFP4 weights on its existing NVIDIA hardware at the reported speed or cost. Those questions require workload-specific measurement and review of the license terms.

Benchmark Against Your Own Tasks

The practical next step for prospective users is to compare providers on the same prompts, context lengths and output requirements they expect in production. Teams should measure end-to-end latency, including reasoning time, and track input and cache costs, since the reported index-task bill is dominated by input. They should also check whether the OpenMDW-1.1 terms fit their intended commercial use and deployment.

Further provider measurements and evaluations could change how Super VL compares on speed and cost. The supplied material does not identify a next benchmark release or a date for additional NVIDIA updates, so the model’s production performance and its standing against newer alternatives remain open questions.

Key Questions

What score did Nemotron 3.5 Super VL receive?

Artificial Analysis reported an Intelligence Index score of 26.5, commonly rounded to 26 in the supplied source. It ranks first among 67 open-weight models in the 40B-to-150B total-parameter class, according to that source.

How fast is the model?

A pre-release DeepInfra endpoint delivered about 630 output tokens per second with 100,000-token prompts. Across six providers, the reported median at that prompt length was 307 tokens per second, with a wide range; speed depends on the endpoint.

How much does an index task cost?

Artificial Analysis estimates $0.25 per Intelligence Index task. That estimate reflects its evaluation workload and pricing assumptions. The source says input represented roughly 89% of the task cost, so costs on other workloads may differ.

Can companies use the model commercially?

The supplied source says the model uses the OpenMDW-1.1 license and allows commercial use. Organizations should review the license terms against their specific deployment plans.

Is it the best-value open model?

Not by the score-and-task-cost comparisons in the source. Several open models score higher at similar or lower reported task costs. Super VL’s case instead includes speed, commercial-use licensing and the option to run released weights on NVIDIA hardware.

Source: ThorstenMeyerAI.com