A Score Of 26: Nemotron 3.5 Super VL’s Performance Snapshot

🔍 Read the full analysis: A Score Of 26: Nemotron 3.5 Super VL’s Performance Snapshot on ThorstenMeyerAI.com

Buying for a business?Offer from Amazon

Get business pricing on monitors, keyboards and dev gear

• 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 26.5 the same day. The 120-billion-parameter open-weight model is designed for text and image input, and its reported combination of speed, price and commercial-use licensing could appeal to teams running workloads on NVIDIA hardware, though its score trails several similarly priced models.

The model has 120 billion total parameters, with 12 billion active in its mixture-of-experts design, and a stated one-million-token context window. NVIDIA released weights in NVFP4, a four-bit format. The source identifies its license as OpenMDW-1.1 and says commercial use is allowed. Artificial Analysis lists API prices of $0.20 per million input tokens and $0.575 per million output tokens, with cached input priced at $0.055 per million tokens.

Artificial Analysis places the model first among 67 open-weight models in its 40-billion-to-150-billion-parameter size class. The source estimates its position at about 18th among open-weight models overall. The distinction matters: the class ranking reflects a size-specific comparison, not leadership across every open-weight model.

For a full Artificial Analysis Intelligence Index task, the reported cost is $0.25, compared with $1.64 for Nemotron 3 Super. The source says the new model scored twice as high as its predecessor, which received 13, while using fewer output tokens per index task: about 46,000 versus 114,000. Input makes up most of the reported bill: the source estimates roughly 89% of the $0.25 cost is input, with cache writes accounting for 69% and output costing less than three cents.

Speed Comes With a Provider Spread

The clearest case for Super VL is operational. Artificial Analysis measured about 630 output tokens per second on a pre-release DeepInfra endpoint using NVFP4 weights and 100,000-token prompts. It described this as the fastest result it had measured at that prompt length among models scoring at least 20. That is a single-endpoint result, though, and does not establish the speed a customer will see elsewhere.

Across six providers, the model page reports a median output speed of 196 tokens per second on its standard prompt and 307 tokens per second with 100,000-token prompts. The measured range runs from 106 to 534 tokens per second. Teams choosing an API provider may therefore see substantial differences and should treat the endpoint they plan to use as the relevant measure.

Median time to the first token is reported at 0.89 seconds. The source says reasoning delays the first answer token to about 11 seconds and a 500-token answer takes about 13.6 seconds end to end. That profile may suit background agents and pipelines that value throughput, while feeling slower in interactive chat. The practical choice depends on the workload, provider and whether users are waiting for each response.

Cost comparisons also put the score in perspective. The source lists MiMo-V2.6-Pro at 46 points for $0.13 per task, MiMo-V2.6-Flash at 38 for $0.06, DeepSeek V4.1 Flash at 39 for $0.27, and GLM-5.3-Flash at 42 for $0.28. Inkling Small scores 26 for $0.09. These figures suggest Super VL is not the cheapest way to reach its score or the highest-scoring model near its task price. Its case instead combines speed, the cited license, NVIDIA provenance and the option of self-hosting.

A Class Ranking Has Limits

The score of 26.5 is the central benchmark figure in the source, which rounds it to 26 in its headline and comparisons. Its reported predecessor, Nemotron 3 Super, scored 13. That makes the new model’s index result twice as high, but the index is an aggregate and does not show performance on every task or deployment.

The same-day comparison includes Muse Spark Open, which the source says scored 39. It describes that model’s Muse Model License as allowing commercial use with restrictions. The source therefore frames Super VL as the leading US option among models with permissive licenses, while noting that the answer changes if a team can accept more restrictive terms.

Some task-level results point to areas of improvement over the previous Nemotron model. The source reports GDPval-AA at 1,162 Elo, up from 490, and AA-Briefcase at 999 Elo, compared with 876 for Nemotron 3 Ultra. It also reports AutomationBench-AA at 14.8%, against 3.0% for Nemotron 3 Ultra. These scores describe performance on named evaluations; they do not establish how the model will perform on a particular company’s software or documents.

Deployment Results Still Vary

The supplied source does not specify the full methodology or uncertainty ranges behind the Intelligence Index and task-level comparisons. The reported 0.25-dollar task cost is tied to Artificial Analysis’s evaluation profile; actual bills will depend on prompt size, caching, output length and provider pricing.

Speed varies by provider, and the 630-token-per-second figure is tied to a pre-release endpoint. The source also gives no independent production results for a specific customer workload. Its hallucination figure falls to 36% on AA-Omniscience, but the source does not define that rate in enough detail here to predict error frequency in every use case. The benchmark does not establish how often errors would occur in a team’s own tasks.

Test the Intended Workload

For teams considering deployment, the reported results point to a practical next step: evaluate the model on the provider and hardware they intend to use, with representative prompts and document sizes. Measure response time, total task cost, answer quality and error rates, including the effect of cached inputs. The source does not announce a future benchmark or product milestone, so any further performance or availability updates remain to be seen.

Teams should also check whether OpenMDW-1.1 meets their licensing and hosting requirements. The reported score and price provide a comparison point; they do not settle whether Super VL is the right choice for a specific production system.

Key Questions

What score did Nemotron 3.5 Super VL receive?

Artificial Analysis reported an Intelligence Index score of 26.5, commonly rounded to 26 in the source. Its predecessor, Nemotron 3 Super, scored 13.

How much does an index task cost?

The source reports a cost of about $0.25 per Artificial Analysis index task. Actual costs depend on token use, caching and provider pricing.

Is it the fastest model in its class?

Artificial Analysis reported about 630 output tokens per second on one pre-release endpoint with 100,000-token prompts. Across six providers, the reported median was 307 tokens per second for that prompt length, so speed depends on the endpoint.

What remains uncertain about production use?

The supplied source does not establish results on specific customer workloads. Teams still need to check performance, costs and error rates with their own prompts, provider and hardware.

Source: ThorstenMeyerAI.com