Nemotron 3.5 Super VL Delivers Sub-Second Latency Per Task

🔍 Read the full analysis: Nemotron 3.5 Super VL Delivers Sub-Second Latency Per Task on ThorstenMeyerAI.com

Before you orderOffer from Amazon

Get smart everyday buys 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, a median first-token time of 0.89 seconds and a $0.25 cost for one index task. The results make it a fast, comparatively inexpensive option, but they do not show it leading on intelligence per dollar: several open-weight models in the source comparison scored higher for similar or lower task costs.

Super VL has 120 billion total parameters, with 12 billion active in its mixture-of-experts design, and accepts text and images while producing text. Its stated context window is one million tokens. NVIDIA released weights in NVFP4, and the model is available through six providers, according to the supplied benchmark material. Its OpenMDW-1.1 license allows commercial use, though prospective users should review the license terms for their own deployment.

Artificial Analysis measured $0.20 per million input tokens and $0.575 per million output tokens, with cached input priced at $0.055 per million tokens. The benchmark’s estimated cost for an Intelligence Index task is $0.25. That is below the reported class median of $0.30 per million input tokens and $0.81 per million output tokens. The source says about 89% of the task cost comes from input, including cache writes, so actual bills will depend on what and how much a system sends to the model.

Speed depends on the measurement. A pre-release DeepInfra endpoint running NVFP4 weights produced about 630 output tokens per second with 100,000-token prompts. Across all six providers, the model page reported a median of 307 tokens per second at that prompt length, with results ranging from 106 to 534. Median time to first token was 0.89 seconds; the source estimates about 11 seconds before the first answer token when reasoning is included, and 13.6 seconds for a 500-token answer.

Speed Comes With a Trade-Off

The results point to a model whose strongest case is operational rather than a clear lead in general capability. Low latency and local deployment can matter for teams building automated pipelines, processing long documents, or needing to run weights on their own NVIDIA hardware. Commercial-use licensing may also fit organizations with requirements that rule out more restrictive model terms.

But task cost is only one part of operating expense, and the sub-second figure describes median time to first token rather than the full time to a reasoned answer. The reported 11-second wait before that answer begins may be acceptable in a background workflow, but noticeable in interactive chat. Provider performance also varies substantially, so a benchmark result from one endpoint is not a guarantee of production speed.

Capability comparisons further qualify the value case. In the supplied figures, MiMo-V2.6-Pro scored 46 for $0.13 per task, while Inkling Small scored 26 for $0.09. Other listed models, including DeepSeek V4.1 Flash and GLM-5.3-Flash, scored above Super VL at roughly comparable task costs. Those comparisons come from the cited benchmark material; they do not establish which model will perform best on a particular company’s workload.

How the Benchmark Frames Its Lead

The score of 26 is double the 13 reported for Nemotron 3 Super, its predecessor. The source also reports the new model’s cost per index task at $0.25, compared with $1.64 for the older model. It attributes the lower cost partly to fewer generated output tokens: about 46,000 per index task, against 114,000 for Nemotron 3 Super.

The headline rank has a defined boundary. Artificial Analysis places Super VL first among 67 open-weight models in its 40-billion-to-150-billion-parameter class. The source estimates it ranks about 18th among open-weight models overall. It also describes it as the top-scoring permissively licensed US open-weight model; that narrower label is not the same as leading every open model or every model available commercially.

The supplied results show improvement in selected tests as well as the overall index. GDPval-AA, a benchmark for economically useful tasks, rose to 1,162 Elo from 490 for Nemotron 3 Super. AutomationBench-AA was 14.8%, compared with 3.0% for Nemotron 3 Ultra. The reported hallucination rate on AA-Omniscience fell from 87% to 36%. These are benchmark-specific measurements, not guarantees of performance on real customer tasks.

Production Results Still Depend on Setup

The available figures do not establish how Super VL will perform across users’ own prompts, hardware, quantization choices or application workflows. The 630-token-per-second result came from one pre-release provider endpoint, while the reported provider range is wide. The source does not provide enough detail here to determine whether every provider used the same hardware and serving configuration.

It is also unclear how closely the benchmark’s estimated $0.25 per task will match production bills. Input volume, cache behavior, reasoning length, provider pricing and repeated requests can all change costs. The listed benchmarks compare selected models, but the material does not supply a full test methodology or independently verify that each model was evaluated under identical settings. Teams should validate relevant tasks and costs themselves before choosing a deployment.

Test Providers on Real Workloads

The immediate next step for prospective users is to test the model through the provider and hardware configuration they expect to use. That means measuring end-to-end response time, not only token throughput, and tracking input, cache and output charges across representative tasks. For teams planning self-hosting, the released NVFP4 weights make local evaluation possible on compatible NVIDIA systems, but the supplied material does not specify the hardware needed for a given throughput target.

Further provider measurements, deployment reports and task-specific evaluations will help establish whether the model’s speed, license and control over weights outweigh its lower general benchmark score relative to some alternatives. Until then, the reported results support a qualified conclusion: Super VL is a low-cost, fast option in its size class, while its practical value depends on the workload and the serving setup.

Key Questions

What is Nemotron 3.5 Super VL?

It is NVIDIA’s 120-billion-parameter open-weight model, with 12 billion active parameters in a mixture-of-experts design. It accepts text and images and generates text.

What does its Intelligence Index score mean?

Artificial Analysis reported a score of 26. In the supplied comparison, that was first among 67 open-weight models in the 40-billion-to-150-billion-parameter class, but the source estimates a rank of about 18th among open-weight models overall.

Does it deliver sub-second answers?

The reported median time to first token was 0.89 seconds. The source estimates that reasoning delays the first answer token to about 11 seconds, so the sub-second metric does not mean a complete answer arrives in under a second.

How much does a task cost?

Artificial Analysis estimated $0.25 per Intelligence Index task. This is a benchmark estimate, not a fixed price for every use; actual charges depend on provider rates, input volume, caching and output.

Is it the best-value open model?

Not on the supplied general benchmark and cost comparisons. Several listed open models scored higher for similar or lower estimated task costs. Super VL’s case instead includes speed, commercial-use licensing and the option to run its weights on local NVIDIA hardware.

Source: ThorstenMeyerAI.com