🔍 Read the full analysis: Nemotron 3.5 Super VL: Measuring AI Speed And Task Cost on ThorstenMeyerAI.com
Get the little things that make your day delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
TL;DR
NVIDIA released Nemotron 3.5 Super VL on Oct. 15, 2026; Artificial Analysis scored it 26, twice its predecessor’s score. It offers low benchmark task costs, fast results on one measured endpoint and commercial use under its OpenMDW-1.1 license, but its speed varies by provider and several competitors score higher for similar or lower costs.
Artificial Analysis lists the model at 26.5 points in its detailed specifications. It ranks first among 67 open-weight models in its 40 billion to 150 billion parameter size class, but around 18th among open-weight models overall. The model has 120 billion total parameters and 12 billion active parameters in a mixture-of-experts design. It accepts text and images, has a stated one-million-token context window, and returns text. NVIDIA released weights in NVFP4, a four-bit format.
The service pricing listed by Artificial Analysis is $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 estimates a full Intelligence Index task costs $0.25. Its launch write-up says Nemotron 3 Super cost $1.64 per task, while output volume fell from about 114,000 tokens per task to about 46,000 for Super VL.
Speed results depend on the measurement. A pre-release DeepInfra endpoint serving NVFP4 weights produced about 630 output tokens per second on prompts of 100,000 tokens. Artificial Analysis describes that as its fastest measured result at that prompt length among models scoring at least 20. Its model page, using the median across six providers, reports 307 tokens per second with 100,000-token prompts and 196 on the standard prompt. The reported 100,000-token speed ranges from 106 tokens per second at the fifth percentile to 534 at the 95th percentile.
Infographic 01 · What 25 cents buys
Cheap and fast. Not the most intelligence per dollar.
Nemotron 3.5 Super VL costs $0.25 per Intelligence Index task and scores 26. Several open models with MIT licenses score 13 to 20 points more for the same money or less.
Intelligence Index vs. cost to run it
Up and to the left is better. The dashed line marks Super VL’s $0.25. Hover a dot for license and values.
Not plotted: K2 Horizon 375B A23B (UAE, index 30.5, listed at $0.00) and Nemotron 3.5 Bolt 74B-A9B (index 21.4, no API price).
License matters here. Muse Spark Open (Meta, 38.9) scores higher than Super VL but ships under the Muse Model License, which allows commercial use with restrictions. Super VL’s OpenMDW-1.1 license allows commercial use.
Show the data as a table (20 models)
| Model | Creator | License | Index | Cost / task |
|---|---|---|---|---|
| MiMo-V2.6-Pro | Xiaomi | MIT | 46.3 | $0.13 |
| GLM-5.3 | Z AI | GLM-5.3 License (restricted) | 44.8 | $2.01 |
| Step 5 | StepFun | StepFun Model License (restricted) | 43.7 | $1.03 |
| Kimi K3 | Kimi | Kimi K3 License (restricted) | 43.6 | $2.00 |
| Claude Haiku 5.5 (max) | Anthropic | Proprietary | 43.4 | $0.21 |
| GLM-5.3-Flash | Z AI | MIT | 41.8 | $0.28 |
| Ling 3.1 Flash | InclusionAI | MIT | 41.1 | $1.00 |
| Qwen3.8 Flash Next | Alibaba | Qwen Community License (restricted) | 39.8 | $0.37 |
| DeepSeek V4.1 Flash | DeepSeek | MIT | 39.5 | $0.27 |
| Muse Spark Open | Meta | Muse Model License (restricted) | 38.9 | $0.84 |
| GPT-6 Luna (max) | OpenAI | Proprietary | 38.1 | $0.07 |
| MiMo-V2.6-Flash | Xiaomi | MIT | 37.9 | $0.06 |
| DeepSeek V4 Pro | DeepSeek | MIT | 36.0 | $0.67 |
| Qwen3.8 27B | Alibaba | Apache 2.0 | 33.7 | $1.01 |
| MiniMax-M3 | MiniMax | MiniMax Community License (restricted) | 29.2 | $0.51 |
| Nemotron 3.5 Super VL | NVIDIA | OpenMDW-1.1 | 26.5 | $0.25 |
| Inkling Small | Thinking Machines | Apache 2.0 | 25.7 | $0.09 |
| Nemotron 3 Ultra 550B | NVIDIA | OpenMDW | 22.9 | $0.60 |
| Muse Glimmer | Meta | Apache 2.0 | 17.5 | $0.05 |
| Nemotron 3 Super (previous) | NVIDIA | NVIDIA Open Model License | 12.8 | $1.64 |
Source: Artificial Analysis, Nemotron 3.5 Super VL model page and model data for the comparison models, read 16 Oct 2026; launch analysis (15 Oct 2026). Intelligence Index v4.3.2, ten evaluations. Cost per task = total API cost to run the Intelligence Index divided by its tasks. “Restricted” = Artificial Analysis lists commercial use as allowed with conditions.
Powered by Thorsten Meyer AI
Speed Depends on the Provider
The gap between the 630-token-per-second DeepInfra result and the 307-token-per-second median across providers changes what operators can expect from the model. A deployment decision based on the fastest published endpoint may not reflect the speed available through another provider. Artificial Analysis’ figures point to provider choice as a major part of the performance picture, so teams evaluating throughput need to measure their intended endpoint.
Latency also affects how the model fits a product. Artificial Analysis reports a median time to first token of 0.89 seconds, compared with 1.76 seconds for Inkling Small and 1.23 seconds for Nemotron 3 Ultra. But the model reasons before answering: the first answer token takes about 11 seconds, and a 500-token response takes around 13.6 seconds end to end. That profile may fit background processing and agent pipelines better than a chat interface where users expect an immediate answer.
The $0.25 task estimate also hides where the bill comes from. Artificial Analysis’ breakdown puts about 89% of task cost on input, with cache writes alone accounting for 69%; output costs less than three cents. For agents that repeatedly process large repositories or document collections, input volume and caching may matter more than the advertised output rate.
As an affiliate, we earn on qualifying purchases.
A Bigger Score, With Tradeoffs
The release follows Nemotron 3 Super, whose Intelligence Index score was 13. The new score doubles that result, while Artificial Analysis reports an approximately 85% lower estimated cost per index task. Those benchmark tasks and cost estimates describe a specific evaluation workload; they do not establish the cost or quality of every production use.
The model’s OpenMDW-1.1 license allows commercial use, according to the source material, and its weights can be run on NVIDIA hardware. That combination may matter to organizations that need control over deployment or require a permissive license. NVIDIA’s standing among U.S. permissively licensed open-weight models is narrower than a claim that it leads all open models: the ranking depends on those licensing and geographic criteria.
Artificial Analysis’ per-evaluation results show uneven performance. GDPval-AA, which measures economically useful work, rose to 1,162 Elo from 490 for Nemotron 3 Super, exceeding Nemotron 3 Ultra’s 1,013 and Inkling Small’s 1,124. AA-Briefcase reached 999 Elo versus 876 for Nemotron 3 Ultra. AutomationBench-AA was 14.8%, about five times Nemotron 3 Ultra’s 3.0%, though the absolute score remains limited. On AA-Omniscience, the reported hallucination rate fell from 87% to 36%.
As an affiliate, we earn on qualifying purchases.
Costs and Scores Need Context
The reported price per task is an Artificial Analysis benchmark estimate, not a universal production bill. Actual costs will depend on workload, prompt length, cache behavior, output length and provider. The source material does not specify how each operator’s infrastructure costs would compare with hosted API pricing.
Speed also remains provider-dependent: the reported range across providers is wide, and the 630-token-per-second figure comes from one pre-release endpoint. The source material does not establish whether every provider can serve the NVFP4 weights at comparable speed or whether these measurements will persist after general availability.
Benchmark results do not settle performance across real-world tasks. The supplied figures show strong gains in selected knowledge-work and automation tests, alongside a 36% hallucination rate on AA-Omniscience. How those results translate to a particular application, and how the model compares on tasks outside the reported evaluations, remains uncertain.
As an affiliate, we earn on qualifying purchases.
Test the Intended Deployment
Teams considering Super VL can compare it on their own prompts and workloads, measuring end-to-end latency, task success and input costs on the provider they expect to use. Long-context evaluations should track cache use and repeated reads, since the reported benchmark cost is driven mainly by input.
Operators should also verify that the model’s license fits their intended use and assess whether the reported quality is sufficient for their application. Artificial Analysis’ release-day benchmarks offer an initial comparison, but provider-specific performance and production costs will become clearer as deployments are measured under real workloads.
As an affiliate, we earn on qualifying purchases.
Key Questions
What is Nemotron 3.5 Super VL’s Intelligence Index score?
Artificial Analysis reports a score of 26 in its release coverage and lists 26.5 in the detailed specifications. Its score is double the 13 reported for Nemotron 3 Super.
How fast is Nemotron 3.5 Super VL?
Artificial Analysis measured about 630 output tokens per second on a pre-release DeepInfra endpoint using 100,000-token prompts. Its median across six providers at that prompt length is 307 tokens per second, so results vary by provider.
How much does a benchmark task cost?
Artificial Analysis estimates $0.25 per Intelligence Index task. That is a benchmark estimate; actual costs depend on the prompt, caching, output and provider.
Can companies use Nemotron 3.5 Super VL commercially?
The supplied source says the model uses the OpenMDW-1.1 license, which allows commercial use. Organizations should review the license terms for their specific use.
Is it the best value among open-weight models?
Not by Intelligence Index score per task in the comparisons provided. Several open-weight models score higher for similar or lower estimated task costs, while Super VL’s case rests more on speed, licensing and the option to run it on NVIDIA hardware.
Source: ThorstenMeyerAI.com
Fall Picks
fall essentials
As an affiliate, we earn on qualifying purchases.
