A Quick Guide To Nemotron 3.5 Super VL’s Key Metrics

🔍 Read the full analysis: A Quick Guide To Nemotron 3.5 Super VL’s Key Metrics 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, with Artificial Analysis reporting an Intelligence Index score of 26.5 and a cost of about $0.25 per index task. The 120-billion-parameter model, which activates 12 billion parameters at a time, is aimed at teams seeking fast inference and commercial use under its OpenMDW-1.1 license; the figures also show that its score trails several competing models available at comparable task costs.

Artificial Analysis lists the model at 26.5 points, rounded to 26 in some comparisons. The source says it ranks first among 67 open-weight models in its 40-billion-to-150-billion-parameter size class, but about 18th when compared with open-weight models across all sizes. It accepts text and images, outputs text and has a stated one-million-token context window. NVIDIA released weights in NVFP4, a four-bit format, and the source reports six API providers.

Speed figures vary by measurement. A pre-release DeepInfra endpoint serving NVFP4 weights reportedly produced about 630 output tokens per second on 100,000-token prompts. Across six providers, Artificial Analysis reports a median of 196 tokens per second on its standard prompt and 307 on 100,000-token prompts, with a provider range from 106 to 534. The 630 figure is an endpoint-specific result, not the reported multi-provider median.

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. The source estimates $0.25 for one Intelligence Index task, compared with $1.64 for Nemotron 3 Super. Its launch write-up attributes much of the lower task cost to fewer generated tokens: about 46,000 output tokens per task versus 114,000 for the predecessor. It also reports that input accounts for about 89% of the newer model’s task cost, with cache writes accounting for 69% and output costing less than three cents.

Speed, Cost and Model Choice

The measurements point to a model whose deployment profile may matter as much as its index score. Teams running large-document or code repository agents may value its reported short time to first token, low cached-input price and option to run the released weights on their own NVIDIA hardware. Artificial Analysis puts median time to first token at 0.89 seconds. The source says reasoning delays the first answer token to about 11 seconds, while a 500-token answer takes about 13.6 seconds end to end; that may suit background workflows better than exchanges where users expect an immediate conversational reply.

For buyers comparing capability per dollar, the cited alternatives complicate the case. The source lists MiMo-V2.6-Pro at 46 points and $0.13 per task, MiMo-V2.6-Flash at 38 and $0.06, DeepSeek V4.1 Flash at 39 and $0.27, and GLM-5.3-Flash at 42 and $0.28. Inkling Small is listed at 26 points and $0.09 per task. These comparisons use the source’s reported Artificial Analysis metrics; they do not establish which model performs best for any particular workload.

License terms are another part of the buying decision. The source describes OpenMDW-1.1 as allowing commercial use and says Nemotron 3.5 Super VL leads among permissively licensed US open-weight models. It contrasts this with Meta’s Muse Spark Open, which scored 39 but uses a license with commercial-use restrictions. Organizations still need to review the license terms against their own use and legal requirements.

How the Scores Compare

The release follows Nemotron 3 Super, which the source lists at 13 on the Intelligence Index. The newer model’s 26.5 score is roughly double that predecessor figure, while its reported task cost is substantially lower. Artificial Analysis’s index is one benchmark summary; it does not by itself measure reliability or quality across every production task.

The source reports gains on selected evaluations. GDPval-AA, described as measuring economically useful work, rose to 1,162 Elo from 490 for Nemotron 3 Super, above the cited scores for Nemotron 3 Ultra (1,013) and Inkling Small (1,124). AA-Briefcase is reported at 999 Elo, compared with 876 for Nemotron 3 Ultra. On AutomationBench-AA, the model scored 14.8%, versus 3.0% for Nemotron 3 Ultra. The source characterizes that as about five times higher, while the reported absolute score remains below 15%.

On AA-Omniscience, the source reports a 36% hallucination rate, down from 87% for Nemotron 3 Super. This is a benchmark result, not a guarantee about the rate of errors in a specific deployment. The source material does not provide full testing methodology for every metric, so readers should treat these figures as evaluation results attributed to Artificial Analysis rather than universal predictions.

Performance Depends on Deployment

The reported speed spread across providers is wide, and the fastest figure came from one pre-release endpoint. Production speed may differ with provider, hardware, prompt length and workload. The source recommends measuring the intended endpoint, but it does not report independent production tests across a range of customer deployments.

Other open questions include how the model performs on a team’s own tasks, whether benchmark gains carry over to its image-based workloads, and how frequently it makes errors under actual operating conditions. The source gives selected benchmark results but does not include a complete account of test methods or confidence intervals. Its cost estimates also depend on the evaluated task and input profile; organizations with different caching, prompt and output patterns may see different bills.

Test Before Production Use

Teams evaluating Nemotron 3.5 Super VL can compare it with alternatives on representative prompts, measuring quality, latency and total cost through the provider or hardware configuration they plan to use. For agent workloads, that includes accounting for repeated document reads and cached input, which the source says make up most of the benchmark task cost.

The next useful evidence will be broader production measurements, further benchmark results and clarification of performance across providers. Until those are available, the published figures establish a fast, lower-cost successor with stronger results on several cited evaluations, but they do not settle whether it is the right model for a particular application.

Key Questions

When was Nemotron 3.5 Super VL released?

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

What does its Intelligence Index score mean?

Artificial Analysis reports a score of 26.5, often rounded to 26 in comparisons. The source says this ranks first among 67 open-weight models in its 40-billion-to-150-billion-parameter size class; it is about 18th among open-weight models overall.

How fast is the model?

A pre-release DeepInfra endpoint reportedly reached about 630 output tokens per second on 100,000-token prompts. Artificial Analysis reports a 307-token-per-second median across providers at that prompt length, with results ranging from 106 to 534.

How much does one benchmark task cost?

Artificial Analysis estimates about $0.25 per Intelligence Index task. Actual costs can vary with input and output volumes, caching and provider pricing.

Is Nemotron 3.5 Super VL the best value?

Not on the cited index score and task price alone. The source lists several models with higher scores at similar or lower estimated costs. Nemotron’s reported speed, license and option to run its weights on NVIDIA hardware may still matter to some teams.

Source: ThorstenMeyerAI.com