Nemotron 3.5 Super VL: What Sub-Second Latency Means For AI

🔍 Read the full analysis: Nemotron 3.5 Super VL: What Sub-Second Latency Means For AI on ThorstenMeyerAI.com

Before you orderOffer from Amazon

Get privacy and security 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, with Artificial Analysis reporting an Intelligence Index score of 26, fast inference on some endpoints and a cost of about $0.25 per index task. The model’s main appeal is operational: it accepts text and images, is available with weights for self-hosting, and carries a license that allows commercial use. Its reported speed is not consistent across providers, and the score does not make it a leader in intelligence per dollar.

Artificial Analysis scored Super VL at 26.5, rounded to 26, on its Intelligence Index. That is twice the score of its predecessor, Nemotron 3 Super, at 13. The evaluation places it first among 67 open-weight models in its 40-billion-to-150-billion total-parameter size group, but around 18th when open-weight models of all sizes are counted. The group ranking is not an overall ranking of every model.

The model has 120 billion total parameters and 12 billion active in its mixture-of-experts design, supports a reported one-million-token context window, and takes text and images as input. NVIDIA released weights in NVFP4, a four-bit format. The source reports an OpenMDW-1.1 license that permits commercial use. Listed API prices are $0.20 per million input tokens and $0.575 per million output tokens; cached input is $0.055 per million tokens.

Speed figures depend on measurement conditions. Artificial Analysis recorded about 630 output tokens per second on a pre-release DeepInfra endpoint using NVFP4 weights and 100,000-token prompts. Across six providers, the reported median was 196 tokens per second on a standard prompt and 307 with 100,000-token prompts, with a range from 106 to 534. Median time to the first token was 0.89 seconds; the source says reasoning can take about 11 seconds before the first answer token, while a 500-token response took about 13.6 seconds end to end.

Speed Comes With Provider Variation

For teams running high-volume pipelines, the combination of short first-token latency, self-hostable weights and comparatively low listed prices could matter more than a top score on a general benchmark. NVIDIA’s figures also point to a substantial change from its previous model: Artificial Analysis put the cost of an index task at $0.25, compared with $1.64 for Nemotron 3 Super, while the new model scored twice as high. Those figures are specific to the index’s evaluation workload, not a guaranteed bill for every use case.

Operational performance needs to be checked at the endpoint a team will actually use. The reported 630 tokens-per-second result came from one pre-release provider, while the six-provider median was much lower. The spread—106 to 534 tokens per second—means hosting choice can materially change throughput. And a fast output rate does not remove the delay before a reasoning model begins answering: the source reports about 11 seconds to the first answer token, a potentially noticeable pause in interactive chat.

Price comparisons also depend on the task. Artificial Analysis estimates that roughly 89% of the $0.25 index-task cost was input, with cache writes accounting for 69% and output costing less than three cents. Applications that repeatedly submit long documents or repository contents may therefore see costs driven mainly by input and caching, rather than by output-token rates alone.

How the Benchmarks Compare

The Intelligence Index is only one view of the model. The source reports stronger results in selected evaluations: its GDPval-AA score rose to 1,162 Elo, from 490 for Nemotron 3 Super, exceeding the cited results for Nemotron 3 Ultra and Inkling Small. On AA-Briefcase, Super VL scored 999 Elo, compared with 876 for Ultra. Its AutomationBench-AA result was 14.8%, about five times Ultra’s 3.0%, although the source notes that the absolute result remains low. On AA-Omniscience, the reported hallucination rate fell from 87% to 36% compared with its predecessor.

Those gains do not establish that Super VL is the best-value model overall. The supplied comparison lists MiMo-V2.6-Pro at 46 for $0.13 per task, MiMo-V2.6-Flash at 38 for $0.06, and DeepSeek V4.1 Flash at 39 for $0.27. GLM-5.3-Flash is listed at 42 for $0.28, while Inkling Small matches Super VL’s score of 26 at $0.09. These comparisons use the source’s reported index scores and per-task costs; buyers should check evaluation methods and their own workload before treating them as direct substitutes.

The licensing comparison has a separate qualification. The source describes Super VL as the highest-scoring permissively licensed US open-weight model, not the highest-scoring open model overall. It says Meta’s Muse Spark Open scored 39 but uses a license with commercial restrictions. A team’s licensing needs, deployment options and data-handling requirements may therefore affect its choice independently of benchmark rank.

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

— Thorsten Meyer, ThorstenMeyerAI.com

Results Depend on Workload and Host

The reported speed varies across providers, and the 630-token-per-second figure was measured on a pre-release endpoint. The source does not provide enough detail here to establish how that result compares under identical hardware, serving settings and production conditions. Its six-provider median offers a broader snapshot, but does not predict a particular customer’s latency or throughput.

It is also unclear how the index-task cost translates to individual applications. The figure depends on the evaluation’s input and output volume, caching and provider pricing. The source material does not provide the complete test protocol for every benchmark cited, nor enough information to assess performance across all image, document or real-time use cases. The reported reduction in hallucination rate is a benchmark result, not a guarantee that the model will avoid errors in deployment.

Test the Intended Deployment

Teams considering Super VL should benchmark it on their own prompts, image inputs and document lengths, using the specific provider or NVIDIA hardware they plan to deploy. They should measure time to first answer token, end-to-end response time, throughput and input-token costs separately, then compare those results with alternatives under the same workload. License terms and any provider-specific service conditions also need review before commercial use.

The next useful evidence will come from production-scale measurements and fuller, comparable evaluations across hosting options. Until those are available, the strongest supported case for Super VL is its mix of fast inference on some endpoints, lower reported task cost than its predecessor, and permissive commercial licensing—not a claim that it offers the highest intelligence per dollar.

Key Questions

What is Nemotron 3.5 Super VL?

It is NVIDIA’s open-weight model, released October 15, 2026. It accepts text and images, generates text, and has 120 billion total parameters, with 12 billion active in its mixture-of-experts design.

How fast is it?

A pre-release DeepInfra endpoint produced about 630 output tokens per second with 100,000-token prompts. Across six providers, Artificial Analysis reported a 307-token-per-second median for that prompt length. Actual performance depends on the endpoint and workload.

Does it cost $0.25 for every task?

No. $0.25 is the reported cost of one Artificial Analysis Intelligence Index task. Actual costs vary with input and output volume, caching and provider pricing.

Is it the top open-weight model?

It ranked first in the cited 40-billion-to-150-billion-parameter size group, but the source places it around 18th among open-weight models overall. Other models in the supplied comparisons scored higher on the Intelligence Index.

Can companies use it commercially?

The source identifies its license as OpenMDW-1.1 and says commercial use is allowed. Organizations should review the license text and any separate provider terms for their intended deployment.

Source: ThorstenMeyerAI.com