A Score Milestone For Nemotron 3.5, But An Open-Weight Gap Remains

🔍 Read the full analysis: A Score Milestone For Nemotron 3.5, But An Open-Weight Gap Remains 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 two open-weight models on Oct. 15, and an evaluation published the same day by Artificial Analysis scored its Nemotron 3.5 Super VL at 26 on the Intelligence Index, up from 13 for its predecessor. The result marks a substantial gain for NVIDIA, but it does not put the model at the top of the open-weight rankings: higher scores from Chinese and UAE developers remain ahead.

The two releases are Nemotron 3.5 Super VL and Nemotron 3.5 Bolt 74B-A9B. Both are mixture-of-experts reasoning models that accept text and images and produce text. Artificial Analysis scored Super VL at 26 (26.5 unrounded) and Bolt at 21. Super VL has 120 billion total parameters, with 12 billion active; Bolt has 74 billion total and 9 billion active.

Artificial Analysis compared Super VL with the earlier Nemotron 3 Super, which has the same stated 120-billion-total, 12-billion-active parameter scale. The earlier model scored 13 and used about 114,000 output tokens per index task, while Super VL scored 26 and used about 49,000. The evaluator’s reported cost per index task fell from $1.64 to $0.25. These are benchmark and cost estimates from the evaluation, not a guarantee of results or pricing in every deployment.

The report also highlights Super VL’s NVFP4 version, a four-bit format. Artificial Analysis found its score within 0.1 points of the full-precision BF16 version, with no individual evaluation changing by more than 2.6 points. The model uses the OpenMDW-1.1 license, which the source says permits commercial use, and the listed API prices are $0.20 per million input tokens and $0.575 per million output tokens. Bolt has no public API listed in the source and is described as self-host only; its license status was not listed by Artificial Analysis.

A stronger model at lower benchmark cost

The release gives developers and businesses another option for workloads where they want to use open weights rather than rely solely on a hosted model. Super VL’s score gain, lower reported token use per index task and lower estimated task cost make the update relevant beyond a simple increase in benchmark performance. Its commercial-use license and reported NVFP4 performance may also matter to teams weighing deployment flexibility and hardware needs.

But the result should not be read as NVIDIA taking the overall open-weight lead. Artificial Analysis places Super VL at 18th among the open-weight models it lists, and 11th when the comparison is restricted to permissively licensed models. The source’s comparison indicates that licensing and performance are separate considerations: a model can be commercially usable while still scoring below competitors, and a high-scoring model may carry commercial restrictions.

Bolt offers a different trade-off. Its score of 21 approaches NVIDIA’s larger Nemotron 3 Ultra score of 23 despite having far fewer parameters, but the evaluator reports about 82,000 output tokens per index task for Bolt, compared with 28,000 for Ultra. Because Bolt has no public API price listed, its real-world cost per answer cannot yet be compared from the supplied information.

Where the models sit in the rankings

Artificial Analysis’ Intelligence Index is the basis for the scores discussed here. In the source’s snapshot, China’s Xiaomi MiMo-V2.6-Pro scored 46 and was listed under an MIT license. Several other Chinese models scored between 44 and 45, though the source says they have commercial conditions in their licenses. The UAE’s K2 Horizon 375B A23B scored 31 and was listed as Apache 2.0.

The article’s count of the ten highest-scoring open-weight models says nine came from Chinese labs, with Meta’s Muse Spark Open the remaining model. That Meta model, released on the same day as NVIDIA’s, scored 39 and is substantially larger at 816 billion total parameters and 42 billion active parameters. The source says it uses the Muse Model License, which has commercial restrictions, unlike Super VL’s stated commercial-use terms.

Artificial Analysis calls Super VL “the leading US open weight model,” a description that needs the source’s licensing qualification. In its comparison, Super VL leads among permissively licensed US models, while the broader list includes higher-scoring models from China and the UAE. The rankings and licensing descriptions are a snapshot reported on Oct. 16, 2026, and may change as models and evaluations are added.

“America’s open-weight position has improved, but it has not changed.”

— Thorsten Meyer, ThorstenMeyerAI.com

Pricing and rankings still have limits

The reported scores come from one evaluator’s index and do not establish how each model will perform on every task. The supplied material does not provide enough detail to independently assess the index’s full methodology or how sensitive the rankings are to changes in its tests. Provider performance also varies: the source reports a pre-release DeepInfra measurement of about 630 output tokens per second for Super VL with 100,000-token prompts, while the median across six providers listed on the model page was 196 tokens per second.

Bolt’s cost remains unknown because the source reports no public API and no measured cost per index task. Its license was also not listed by Artificial Analysis. More broadly, the source does not establish whether the rankings will remain the same as more providers, models or evaluations are added, or how the reported benchmark costs translate to a buyer’s specific workload.

Deployment data will test the results

The next practical test is broader availability and use outside the initial evaluation. Super VL’s listed API prices and reported provider speed measurements give prospective users some basis for comparison, but deployment results will depend on provider, hardware, prompt length and task. Teams considering self-hosting will also need to test memory use and output quality with their own workloads rather than treating benchmark results as a substitute.

For Bolt, a public API, a published license and independently measured serving costs would clarify how its smaller active parameter count translates into practical access and expense. Future Artificial Analysis updates could also change the relative rankings. Until then, the confirmed picture is a marked NVIDIA score improvement, alongside a continuing performance gap with higher-ranked open-weight models.

Key Questions

What did NVIDIA release?

NVIDIA released Nemotron 3.5 Super VL and Nemotron 3.5 Bolt 74B-A9B on Oct. 15, 2026. Both accept text and images and return text, according to the source.

How did Super VL score?

Artificial Analysis scored Super VL at 26 on its Intelligence Index, or 26.5 before rounding. The predecessor, Nemotron 3 Super, scored 13 in the comparison reported by the source.

Does Super VL lead all open-weight models?

No. The source places it 18th among open-weight models listed by Artificial Analysis and 11th among permissively licensed models. It describes Super VL as the leading US model among permissively licensed models, while higher-scoring models from China and the UAE remain ahead in the broader comparison.

Can businesses use Super VL commercially?

The source says Super VL uses the OpenMDW-1.1 license, which allows commercial use. Prospective users should check the license terms themselves before deployment; the source describes Bolt’s license as not yet listed by Artificial Analysis.

Is Bolt cheaper to run than larger models?

That is not yet established by the supplied information. Artificial Analysis reports that Bolt used about 82,000 output tokens per index task, and the source says it has no public API or measured task cost. Token use alone does not determine total deployment cost.

Source: ThorstenMeyerAI.com