A Higher Score For Nemotron 3.5, An Unchanged Open-Weight Challenge

🔍 Read the full analysis: A Higher Score For Nemotron 3.5, An Unchanged Open-Weight Challenge 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 and Nemotron 3.5 Bolt 74B-A9B on October 15, with Artificial Analysis scoring Super VL at 26 on its Intelligence Index, twice the 13 recorded for Nemotron 3 Super. The results mark a sizable improvement for NVIDIA’s open-weight models, while the same evaluation places the United States behind Chinese and UAE models in the wider field.

Artificial Analysis reports Super VL’s unrounded score as 26.5. The model has 120 billion total parameters, of which 12 billion are active per token, and accepts text and images. It uses the OpenMDW-1.1 license, which permits commercial use, and has a one-million-token context window. Artificial Analysis lists API prices of $0.20 per million input tokens and $0.575 per million output tokens.

The comparison with Nemotron 3 Super is notable on more than the index score. Artificial Analysis estimates output fell from about 114,000 tokens to 49,000 per index task, while task cost fell from $1.64 to $0.25. The source describes the models as the same size class. These are evaluation-specific measurements; actual costs and speeds depend on the task and provider.

NVIDIA’s second model, Bolt 74B-A9B, scored 21, with 74 billion total parameters and nine billion active. It has a 256,000-token context window and accepts text and images. Artificial Analysis says Bolt has no public API yet and lists no license for it on the model page, so its hosted price and license terms remain unavailable there.

A Better Model, a Wider Gap

Super VL’s score and task-cost figures suggest a more capable model that can also use fewer output tokens in the benchmark’s tasks. That combination may matter to developers choosing a model for self-hosting or paid API use. Artificial Analysis also says the model’s NVFP4 4-bit weights scored within 0.1 points of full-precision BF16 weights overall, with no individual evaluation moving by more than 2.6 points. That result points to a way to reduce memory needs, though deployment results can vary by hardware and setup.

The international ranking tempers the significance of NVIDIA’s progress. The source says Super VL ranks 18th among listed open-weight models, or 11th when restricted to permissively licensed models. It leads the permissively licensed US models in the cited comparison, but remains well below the leading Chinese model. The release improves NVIDIA’s position without, on these rankings, changing the United States’ broader standing.

License terms shape who can use a model commercially. The source says Meta’s Muse Spark Open scored 39 on the same day, but uses a license with commercial restrictions. That makes direct comparisons of score alone incomplete for organizations weighing adoption; the permitted uses and deployment costs also matter.

How the Models Compare

Artificial Analysis evaluated the two models on October 15, the day NVIDIA released them. Both are mixture-of-experts reasoning models: only some of their parameters are active for each token. Super VL is the flagship, while Bolt begins a separate NVIDIA model family aimed at smaller deployment requirements.

The source reports Super VL at $0.25 per index task, compared with $0.60 for NVIDIA’s earlier Nemotron 3 Ultra, which scored 23 and has 550 billion total parameters with 55 billion active. Bolt scored 21, 1.5 points below Ultra, despite having much smaller parameter counts. It performed better than Ultra on two cited evaluations, AutomationBench-AA and AA-Briefcase, and lower on long-context reasoning and Humanity’s Last Exam.

Those results come with an efficiency caveat: the source estimates Bolt generates about 82,000 output tokens per index task, versus about 28,000 for Ultra. Because Bolt has no public API, there is no listed price per task to show whether its lower per-token operating needs translate into cheaper answers.

Rankings and Pricing Still Shift

The supplied source does not include the complete license terms for Meta’s Muse Model License, so the specific commercial restrictions cannot be detailed here. It also says Artificial Analysis had not listed a license for Bolt and that the model had no public API; both its permitted uses and hosted pricing therefore remain unclear from the cited material.

Rankings reflect the models Artificial Analysis listed at the time of its evaluation and may change as new models are added or existing scores are updated. The source reports rankings and selected benchmark results but does not provide enough detail here to independently assess every model’s evaluation conditions. Speed is also provider-dependent: a pre-release DeepInfra endpoint reached about 630 output tokens per second on 100,000-token prompts, while the median across six providers listed on the model page was 196 tokens per second.

Watch for Bolt Access and Terms

The next practical developments are public access to Bolt, publication of its license and pricing, and further provider measurements for Super VL. Those details will help users compare deployment costs and performance beyond the initial benchmark results. Artificial Analysis may also update its rankings as new evaluations appear; the supplied source does not give a date for any such update.

Key Questions

What did NVIDIA release?

NVIDIA released Nemotron 3.5 Super VL and Nemotron 3.5 Bolt 74B-A9B on October 15, 2026. Both accept text and images and return text.

How did Super VL score?

Artificial Analysis gave Super VL an Intelligence Index score of 26, or 26.5 unrounded. Its predecessor, Nemotron 3 Super, scored 13.

Does Super VL allow commercial use?

The source identifies Super VL’s license as OpenMDW-1.1 and says it allows commercial use.

Is Super VL the top-ranked open-weight model?

No. The source places it 18th among listed open-weight models and 11th among permissively licensed models. It leads the permissively licensed US models in the cited comparison, while models from China and the UAE score higher overall.

Can users compare Bolt’s price yet?

Not from the supplied figures. Artificial Analysis lists no public API for Bolt and says no price per index task has been measured.

Source: ThorstenMeyerAI.com