Nemotron 3.5 Super VL: A Score Of 26 At Sub-Second Speed

🔍 Read the full analysis: Nemotron 3.5 Super VL: A Score Of 26 At Sub-Second Speed on ThorstenMeyerAI.com

Buying for a business?Offer from Amazon

Get business pricing on office and shipping supplies

• 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 rated it 26 on its Intelligence Index, twice the score of its predecessor, Nemotron 3 Super. The evaluation also found high throughput on one pre-release endpoint; the model’s median speed across six providers was lower, a distinction that matters to teams weighing deployment costs and response times.

Artificial Analysis lists the model at 26.5 points and ranks it first among 67 open-weight models in its 40-billion-to-150-billion-parameter class. Its approximate position across all open-weight models is 18th. The model has 120 billion total parameters, with 12 billion active in its mixture-of-experts design, and a stated context window of one million tokens. It accepts text and images and generates text.

The listed API rates are $0.20 per million input tokens and $0.575 per million output tokens, with cached input at $0.055 per million. Artificial Analysis estimates a full Intelligence Index task costs $0.25. That compares with $1.64 for Nemotron 3 Super, which scored 13. The source attributes the lower task cost in part to fewer generated tokens: about 46,000 per index task, compared with 114,000 for the predecessor.

Speed depends on the measurement. A pre-release DeepInfra endpoint serving NVFP4 weights produced about 630 output tokens per second with 100,000-token prompts. Artificial Analysis described that as its fastest measurement at that prompt length among models scoring at least 20. Across six providers, however, the median was 196 tokens per second on standard prompts and 307 with 100,000-token prompts; observed output rates ranged from 106 to 534 tokens per second.

Speed and Cost for Operators

The release presents a trade-off for teams choosing models to run in production. Lower estimated task costs and a commercial-use license may matter to organizations that want to host the model on NVIDIA hardware or use available providers. The source describes its OpenMDW-1.1 license as permitting commercial use, and says the weights are released in NVFP4.

For interactive applications, throughput alone does not describe the user experience. Artificial Analysis reports a median time to first token of 0.89 seconds, but says the model takes about 11 seconds to produce its first answer token because it reasons before answering. A 500-token answer takes about 13.6 seconds end to end. That timing may suit background workflows better than chat interfaces where users expect an immediate response.

The task-cost estimate also depends on what a system sends to the model. The source says about 89% of the estimated $0.25 task cost is input, with cache writes accounting for 69% and output costing less than three cents. For agents that repeatedly process large repositories or document sets, actual spend will depend on input volume and caching patterns, not just the advertised output rate.

How the Score Compares

The 26-point score is a marked improvement over Nemotron 3 Super’s 13, but does not make Super VL the strongest value by the source’s comparisons. It 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 also listed at 26, with an estimated task cost of $0.09.

Those figures compare benchmark scores and estimated costs; they do not establish that models perform equally well on every workload or have equivalent deployment characteristics. The source also compares Super VL with proprietary tiers, including Claude Haiku 5.5 at 43 points and $0.21 per task, and GPT-6 Luna at 38 and $0.07. Teams may weigh those comparisons alongside hosting, licensing and endpoint availability.

The claim of a leading US model is narrower than a general ranking. The source says Super VL has the top score among permissively licensed US open-weight models. It contrasts that category with Meta’s Muse Spark Open, which scored 39 but uses a license the source describes as commercially restricted. The category depends on how a team interprets its licensing requirements.

Limits of the Available Measurements

The reported figures do not settle how the model will perform on every production workload. Provider speeds vary substantially, and the 630-token-per-second result came from one pre-release endpoint. The source advises measuring the endpoint a team plans to use. The available material does not specify whether those provider measurements will remain stable after launch or how performance changes with different hardware and request patterns.

The source also reports improvements in particular evaluations: GDPval-AA rose to 1,162 Elo from 490 for the predecessor, while AutomationBench-AA reached 14.8%, compared with 3.0% for Nemotron 3 Ultra. It says the AA-Omniscience hallucination rate fell from 87% to 36%. These are benchmark results, not guarantees about a deployed system. The provided material does not give enough detail to assess every benchmark’s methodology or predict performance on a reader’s own tasks.

Task costs are estimates tied to Artificial Analysis’s evaluation profile and listed rates. Actual bills will depend on token use, cache behavior, provider pricing and workload. The source does not provide a full account of hardware costs for self-hosting, so the API task estimate should not be treated as the total cost of operating the model.

Test the Intended Deployment

For teams considering Super VL, the next step is to measure it on their own prompts and chosen provider. Record response latency, output speed, task accuracy and token use, including how often requests reuse cached input. Those results can show whether the model’s price and throughput fit a particular workload.

Organizations planning self-hosting will also need to evaluate their available NVIDIA hardware and the released NVFP4 weights. The source identifies commercial use as allowed under OpenMDW-1.1, but teams should check the license terms against their own requirements. The supplied reporting gives no later launch milestone or scheduled update; real-world operating costs and performance remain to be established by deployment.

Key Questions

What score did Nemotron 3.5 Super VL receive?

Artificial Analysis lists a score of 26.5 on its Intelligence Index, reported in the source as 26. It also ranks first among 67 open-weight models in the 40-billion-to-150-billion-parameter class.

Is 630 tokens per second the speed most users should expect?

No single speed is established for every provider. The 630-token-per-second figure came from one pre-release DeepInfra endpoint. Across six providers, Artificial Analysis reported median output speeds of 196 tokens per second on standard prompts and 307 with 100,000-token prompts.

How much does an Intelligence Index task cost?

Artificial Analysis estimates $0.25 per task for Super VL. The source says about 89% of that amount is input cost; actual spending depends on token use, caching and provider rates.

Can businesses use the model commercially?

The source identifies the OpenMDW-1.1 license and says it allows commercial use. Organizations should review the license terms for their intended deployment.

Source: ThorstenMeyerAI.com