Under A Second Per Task: Nemotron 3.5 Super VL’s Latency

🔍 Read the full analysis: Under A Second Per Task: Nemotron 3.5 Super VL’s Latency 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, and Artificial Analysis reported an Intelligence Index score of 26, with a median first-token latency of 0.89 seconds and an estimated cost of $0.25 per benchmark task. The figures position it as a fast, relatively inexpensive open-weight model, while showing that several similarly priced models score higher on the index.

The model has 120 billion total parameters, with 12 billion active in its mixture-of-experts architecture, and supports a reported one-million-token context window. Its released weights use NVFP4, and its OpenMDW-1.1 license allows commercial use, according to the source material. 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.

Speed depends heavily on the provider. A pre-release DeepInfra endpoint serving NVFP4 weights produced about 630 output tokens per second on 100,000-token prompts, a result Artificial Analysis described as its fastest measurement at that prompt length among models scoring at least 20. Across six providers, however, the model page reports a median of 307 output tokens per second on those long prompts, ranging from 106 at the fifth percentile to 534 at the 95th percentile. Its median output speed on standard prompts is 196 tokens per second.

Artificial Analysis reports a $0.25 cost per Intelligence Index task, compared with $1.64 for Nemotron 3 Super. The new model’s score is twice its predecessor’s 13, while its reported output volume per index task fell from about 114,000 tokens to 46,000. Input accounts for roughly 89% of the new model’s task cost in the published breakdown, including 69% from cache writes. That makes actual spending dependent on how much input an application sends and caches.

Fast Responses, Uneven Value

The results matter to teams weighing response time, operating cost and model quality for production systems. A median 0.89-second wait for the first token may suit interactive applications, while lower per-task costs and the option to run released weights on NVIDIA hardware can appeal to organizations with infrastructure or deployment requirements of their own.

But latency is not the same as time to a useful answer. The source reports that the model reasons for about 11 seconds before its first answer token; a 500-token answer takes about 13.6 seconds end to end. That delay may be acceptable for background agents and automated pipelines but noticeable in a live chat. The source also says MiMo-V2.6-Pro scored 46 at $0.13 per task, while Inkling Small scored 26 at $0.09. On the reported index and task cost, Super VL is not the strongest value for raw benchmark performance.

Its case instead depends on factors beyond that score: speed on a chosen endpoint, commercial licensing, model provenance and self-hosting. The published comparison says it leads among permissively licensed US open-weight models. That distinction may matter to procurement and legal teams, though the source notes that other models have different licensing terms and restrictions.

How the Benchmark Compares

Artificial Analysis assigns Nemotron 3.5 Super VL an Intelligence Index score of 26 (reported as 26.5 in the specification table). The source says it ranks first among 67 open-weight models in its 40-billion-to-150-billion-total-parameter size class. Across open-weight models overall, it places around 18th, according to the source. The size-class ranking therefore describes a narrower comparison than a claim of overall leadership.

Other reported results point to specific improvements over Nemotron 3 Super. On GDPval-AA, a benchmark for economically useful tasks, Super VL scored 1,162 Elo, up from 490; it also scored 999 Elo on AA-Briefcase, compared with 876 for Nemotron 3 Ultra. AutomationBench-AA was 14.8%, versus 3.0% for Nemotron 3 Ultra. The source cautions that the automation result remains low in absolute terms. On AA-Omniscience, it reports a hallucination rate of 36%, down from 87% for the predecessor.

The comparisons cited in the source include MiMo-V2.6-Flash at 38 for $0.06 per task, DeepSeek V4.1 Flash at 39 for $0.27, and GLM-5.3-Flash at 42 for $0.28. These are benchmark and cost figures reported by the source; they do not establish how each model will perform on a particular team’s workload. The source also says Meta’s Muse Spark Open scored 39 but uses a license with commercial restrictions.

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

— Thorsten Meyer, ThorstenMeyerAI.com

Provider Results May Vary

The reported 630-token-per-second result came from one pre-release endpoint, while the provider-wide median was lower. The source does not establish whether production deployments will reproduce that peak, and says provider choice can change observed speed by a factor of two to five. Buyers will need measurements from the specific endpoint and workload they intend to use.

Benchmark scores and estimated task costs are not guarantees of performance or bills for every application. The source does not specify how the model performs across all real-world use cases, nor does it provide enough detail here to compare reliability, hardware requirements or operating costs for self-hosted deployments. Its reported hallucination rate also comes from a particular benchmark and should not be read as a universal rate.

Measure the Intended Workload

Teams considering deployment can compare the model’s latency, answer quality and input-token costs on their own tasks and provider. That is especially relevant for applications that repeatedly process long documents or repositories, since the source’s cost breakdown attributes most of the benchmark-task bill to input and cache writes.

Further provider measurements and independent evaluations could clarify how stable the speed results are and whether the benchmark gains translate to production workloads. The source material gives no scheduled next evaluation or deployment milestone.

Key Questions

When was Nemotron 3.5 Super VL released?

NVIDIA released Nemotron 3.5 Super VL on October 15, 2026. Artificial Analysis published figures the same day, according to ThorstenMeyerAI.com.

How fast is the model?

Artificial Analysis reports median output speeds of 196 tokens per second on standard prompts and 307 tokens per second on 100,000-token prompts across six providers. A pre-release DeepInfra endpoint reached about 630 tokens per second on long prompts; that was a single-endpoint result.

How much does an Intelligence Index task cost?

The reported estimate is $0.25 per task. The source’s cost breakdown says about 89% comes from input, including cache writes, so actual costs depend on an application’s input and caching patterns.

Does it lead all open-weight models?

No. The source reports a score of 26 and a first-place ranking among 67 open-weight models in its 40-billion-to-150-billion-parameter size class. It places around 18th among open-weight models overall, and several named alternatives have higher scores.

What remains unknown about its production performance?

The cited speed peak came from one pre-release endpoint, and the source does not establish that other providers or production deployments will match it. Results on a team’s own workload, along with self-hosting requirements and costs, also need separate evaluation.

Source: ThorstenMeyerAI.com