25 Cents A Task: Exploring Nemotron 3.5 Super VL

🔍 Read the full analysis: 25 Cents A Task: Exploring Nemotron 3.5 Super VL on ThorstenMeyerAI.com

Before you orderOffer from Amazon

Get the latest gadgets 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, 2026, and Artificial Analysis reported an Intelligence Index score of 26.5 and a cost of about $0.25 per index task. The model combines open weights, commercial use under the OpenMDW-1.1 license and high reported throughput, but its benchmark score trails several alternatives at similar or lower task costs.

The model has 120 billion total parameters, with 12 billion active in its mixture-of-experts design. It accepts text and images, produces text, and has a stated context window of one million tokens. NVIDIA released weights in NVFP4, a four-bit format. The source says the license allows commercial use, though teams should review its terms for their intended deployment.

Artificial Analysis places the model first among 67 open-weight models in its 40-billion-to-150-billion-parameter size class. That ranking is not an overall open-model lead: counting each model once, the source puts it at about 18th among open-weight models. Its reported 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 depends heavily on the measurement. A pre-release DeepInfra endpoint serving NVFP4 weights produced about 630 output tokens per second on prompts of 100,000 tokens. Across six providers, Artificial Analysis lists a median of 307 tokens per second at that prompt length, with results ranging from 106 to 534. Median time to the first token was 0.89 seconds, but the source reports about 11 seconds before the first answer token after reasoning, and 13.6 seconds for a 500-token answer.

Where Speed Meets Model Value

The release matters to teams weighing throughput, deployment control and cost against answer quality. At $0.25 per Intelligence Index task, the model is less expensive than its predecessor, Nemotron 3 Super, which the source puts at $1.64 per task. Its score also doubled from 13 to 26.5, while output volume per task fell from about 114,000 tokens to 46,000, according to the source’s account of Artificial Analysis figures.

That improvement does not make Super VL the strongest value on the cited benchmark. The source lists MiMo-V2.6-Pro at 46 for $0.13 per task, MiMo-V2.6-Flash at 38 for $0.06, and Inkling Small at 26 for $0.09. The comparison suggests the NVIDIA model’s value depends on requirements beyond benchmark score and task price. Teams may place greater weight on self-hosting, the stated license, latency or compatibility with NVIDIA hardware.

The price breakdown also matters for applications that repeatedly process long inputs. The source estimates about 89% of the task cost comes from input, with cache writes alone accounting for 69%, while output costs less than three cents. Actual bills will depend on prompt length, caching, provider and workload, so the headline task cost should not be treated as a universal per-request price.

How the Benchmark Comparisons Work

The Intelligence Index is a benchmark score, not a direct measure of performance in every company’s product. Artificial Analysis compares models, and the source’s claim that Super VL ranks first applies specifically to the 40B–150B open-weight size class. Its broader rank is lower. That distinction is key when interpreting a headline ranking.

The speed figures also describe different test conditions. The approximately 630-token-per-second result came from one pre-release endpoint; the multi-provider figures use medians across six providers. A business using a different provider or deployment may see materially different throughput. The source recommends testing the endpoint a team plans to use rather than relying on the single fastest result.

The source also reports substantial gains on selected evaluations: GDPval-AA, which covers economically useful tasks, rose to 1,162 Elo from 490 for Nemotron 3 Super, while AA-Briefcase reached 999 Elo, up from 876 for Nemotron 3 Ultra. AutomationBench-AA was 14.8%, compared with 3.0% for Ultra. These are benchmark results reported in the source, not guarantees of performance on a particular workflow.

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

— Thorsten Meyer, ThorstenMeyerAI.com

What Deployment Results May Vary

The reported speed figures are not a single consistent result: the fastest number came from a pre-release DeepInfra endpoint, while provider-wide medians were lower. The source does not provide enough detail here to establish how every provider’s hardware, serving configuration or workload compares, so production performance remains uncertain until tested on the intended setup.

The source also does not establish how well the benchmark results transfer to any particular application. Although it reports a lower hallucination rate on AA-Omniscience—36%, down from 87% for Nemotron 3 Super—the figures are evaluation results, not proof that errors will occur at the same rate in customer-facing use. The excerpt gives no detailed breakdown of the measure or its task conditions.

It is also unclear from the supplied material how NVIDIA’s license terms apply to every commercial scenario, or what self-hosting will cost after hardware, engineering and maintenance are included. The cited API price and task-cost estimate do not answer those deployment questions.

Test Costs Before Production

Teams considering Nemotron 3.5 Super VL should benchmark it on their own prompts, image workloads and provider endpoint, tracking response quality, first-answer latency, throughput and input-token use. Long-context applications should measure cache behavior directly because input and cache writes account for most of the source’s estimated task cost.

Before deployment, operators will also need to check the OpenMDW-1.1 license against their use case and compare API use with self-hosting costs on suitable NVIDIA hardware. The supplied source does not give a date for further releases or independent production evaluations. Provider performance, pricing and outside evaluations may change, so comparisons should be revisited as newer data becomes available.

Key Questions

When was Nemotron 3.5 Super VL released?

NVIDIA released it on October 15, 2026. Artificial Analysis published benchmark figures the same day, according to the source.

What does the Intelligence Index score mean?

The reported score is 26.5. The source says the model ranks first among 67 open-weight models in the 40B–150B parameter size class, but about 18th among open-weight models overall. The score does not predict results for every application.

How much does one task cost?

Artificial Analysis lists about $0.25 per Intelligence Index task. The source estimates that most of this cost comes from input and cache writes, so an individual workload’s bill will vary with its token use and caching.

Is the model fast in production?

One pre-release endpoint produced about 630 output tokens per second on 100,000-token prompts, while the median across six providers was 307 tokens per second for that prompt length. Results vary by provider, and the source reports about 11 seconds before the first answer token after reasoning.

Can businesses use the model commercially?

The source identifies its license as OpenMDW-1.1 and says commercial use is allowed. Businesses should review the license terms for their specific deployment before relying on that summary.

Source: ThorstenMeyerAI.com