Nemotron 3.5 Super VL: An AI Model With Fast, Low-Cost Tasks

🔍 Read the full analysis: Nemotron 3.5 Super VL: An AI Model With Fast, Low-Cost Tasks 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 scored it 26 on its Intelligence Index. The model combines low reported evaluation costs with high throughput on one pre-release endpoint, but comparisons in the launch material show several alternatives scoring higher for similar or less money, a distinction that matters to teams choosing a model for production.

The model has 120 billion total parameters, with 12 billion active in its mixture-of-experts design, and accepts text and images. NVIDIA released its weights in NVFP4 format; the source describes its OpenMDW-1.1 license as allowing commercial use. The stated context window is one million tokens. 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 tokens.

Artificial Analysis reports a cost of $0.25 per Intelligence Index task, compared with $1.64 for Nemotron 3 Super. The score rose from 13 for that predecessor to 26 for Super VL. In the launch write-up, Super VL used about 46,000 output tokens per index task, against 114,000 for its predecessor. The source’s cost breakdown says about 89% of the $0.25 came from input, including 69% from cache writes; output cost less than three cents.

Speed varies by measurement. A pre-release DeepInfra endpoint serving the NVFP4 weights produced about 630 output tokens per second on prompts of 100,000 tokens. Artificial Analysis describes that as the fastest model it had measured at that prompt length among models scoring at least 20. Its model page, using the median across six providers, instead lists 196 tokens per second on a standard prompt and 307 on 100,000-token prompts, with provider results ranging from 106 to 534 tokens per second. Median time to first token is 0.89 seconds; the source says reasoning takes about 11 seconds before the first answer token, and a 500-token response takes about 13.6 seconds end to end.

Cost and Speed Versus Model Quality

For operators, the figures point to a model that may suit high-volume pipelines and background agents where throughput, local deployment options, and predictable input costs matter. The source says a 500-token answer takes around 13.6 seconds end to end, so the reported speed advantage does not automatically translate into a fast-feeling conversational product. Teams should also expect performance to depend on the provider: the reported output rate spans more than a fivefold range across the six listed providers.

The score and cost comparisons complicate a simple value claim. The source lists MiMo-V2.6-Pro at 46 points for $0.13 per task, MiMo-V2.6-Flash at 38 for $0.06, and Inkling Small at 26 for $0.09. It also cites DeepSeek V4.1 Flash at 39 for $0.27 and GLM-5.3-Flash at 42 for $0.28. These are figures reported in the source material, not a guarantee that models will perform equally across a buyer’s own workloads. They suggest Super VL’s strongest case is its combination of speed, licensing, and self-hosting, rather than the highest score per dollar.

Licensing may be decisive for some organizations. The source characterizes Super VL as the leading option among permissively licensed US open-weight models, while noting that Meta’s Muse Spark Open scored 39 but uses a license with commercial-use restrictions. That comparison depends on a team’s legal requirements and on the source’s definition of permissive licensing; it does not establish that Super VL is the only suitable model for every company.

How Super VL Compares

Artificial Analysis places Super VL first among 67 open-weight models in its 40-billion-to-150-billion-parameter size class. That category ranking has a narrower scope than the full open-weight field: the source places Super VL around 18th when all open-weight models are counted. Its score of 26 is therefore a marked improvement over Nemotron 3 Super, but it does not put it among the highest-scoring open models overall.

The launch material also reports uneven benchmark results. On GDPval-AA, which evaluates economically useful tasks, Super VL reached 1,162 Elo, up from 490 for Nemotron 3 Super and above the cited scores for Nemotron 3 Ultra and Inkling Small. Its AA-Briefcase score was 999 Elo, compared with 876 for Nemotron 3 Ultra. AutomationBench-AA was 14.8%, against 3.0% for Ultra. The source calls that roughly five times higher, while also describing the absolute result as low.

One reported change concerns hallucinations: the AA-Omniscience hallucination rate fell from 87% to 36% compared with Nemotron 3 Super. That is a substantial improvement on the cited measure, but a 36% rate still indicates a meaningful limitation. Benchmark outcomes describe specific tests; they do not by themselves establish accuracy on an organization’s documents or workflows.

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

— Thorsten Meyer, ThorstenMeyerAI.com

Provider Results and Workload Fit

The launch figures do not establish what speed or cost every customer will see. The 630-token-per-second result came from a single pre-release endpoint, while provider medians and ranges show substantial variation. The source advises measuring the endpoint a team plans to use; actual results may also depend on prompt length, caching, output length, and workload.

The available material does not specify how Super VL performs across a broad range of image tasks, nor does it provide enough detail to determine whether the benchmarks predict performance for a particular production use. The meaning of the reported one-million-token context window for practical accuracy and cost is also not established here. The license’s terms and fit for any particular deployment should be checked directly against the license text.

Test the Intended Deployment

Teams considering the model should test their own prompts and documents using the provider and hardware they would deploy. That can establish response time, token use, cache behavior, accuracy, and failure rates for their workload. The source lists six providers, but does not identify their individual terms or say whether the pre-release DeepInfra result will be available in the same form after launch.

Further provider measurements and independent evaluations could clarify how stable the speed and cost figures are outside the launch tests. Until then, the reported results support a focused conclusion: Super VL offers a notable score improvement over its predecessor and low task costs, while its fit depends on whether buyers value its licensing and deployment options enough to accept a lower overall intelligence score than several cited alternatives.

Key Questions

What score did Nemotron 3.5 Super VL receive?

Artificial Analysis gave it an Intelligence Index score of 26 (the source also lists 26.5). Its predecessor, Nemotron 3 Super, scored 13.

How fast is Nemotron 3.5 Super VL?

A pre-release DeepInfra endpoint produced about 630 output tokens per second on 100,000-token prompts. Artificial Analysis lists a six-provider median of 307 tokens per second on prompts of that length, with results ranging from 106 to 534.

How much does an Intelligence Index task cost?

Artificial Analysis reports a cost of $0.25 per task. This is an evaluation cost; the cost of a specific production workload will depend on its input, output, and caching profile.

Can companies use the model commercially?

The source says the model uses the OpenMDW-1.1 license and allows commercial use. Organizations should review the license itself to assess its terms for their intended deployment.

Is it the best-value open-weight model?

Not by the score-and-cost comparisons in the source. Several named open-weight alternatives scored higher for similar or lower reported task costs. Super VL’s case also includes speed, licensing, provenance, and the option to run it on NVIDIA hardware.

Source: ThorstenMeyerAI.com