The Nemotron 3.5 Score Gain Doesn’t Erase America’s Open-Weight Gap

🔍 Read the full analysis: The Nemotron 3.5 Score Gain Doesn’t Erase America’s Open-Weight Gap on ThorstenMeyerAI.com

Before you orderOffer from Amazon

Get smart everyday buys delivered free with Prime

• Fast, free delivery on millions of items
• Prime Video, Amazon Music and more included
• Member-only deals all year

Start your free Prime trial →Free trial for eligible customers · Cancel anytime

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 same-day Artificial Analysis results showing improved performance from the company’s previous generation. Super VL scored 26 on the Intelligence Index, twice the 13 recorded for Nemotron 3 Super, but it remains behind top-ranked models from China and the UAE in the rankings described by the source.

Artificial Analysis lists Super VL at 26 points (26.5 before rounding) and Bolt at 21 points (21.4 unrounded). Both models accept text and images and produce text. They use a mixture-of-experts design, activating only part of their networks for each token. Super VL has 120 billion total parameters, with 12 billion active, and a one-million-token context window. Bolt has 74 billion total parameters, nine billion active, and a 256,000-token context window.

The evaluation reports a marked efficiency change for Super VL compared with its predecessor: output tokens per index task fell from about 114,000 to 49,000, while the reported cost per task dropped from $1.64 to $0.25. Artificial Analysis also found the model’s NVFP4 four-bit version scored within 0.1 points of full-precision BF16 weights. Those results concern the tested model and evaluation; performance and serving costs can vary by provider and use case.

Bolt’s score is close to the 23 recorded for NVIDIA’s larger Nemotron 3 Ultra, despite Bolt having far fewer parameters. However, Artificial Analysis measured about 82,000 output tokens per index task for Bolt, compared with 28,000 for Ultra, and the source says Bolt had no public API at the time of its report. Its cost per task therefore had not been measured.

Efficiency Gains, Ranking Limits

The release matters for developers choosing models that can be adapted, inspected or run on their own infrastructure. Super VL combines a higher benchmark score with lower reported token use and task cost than its predecessor. Its OpenMDW-1.1 license allows commercial use, according to the source, while its NVFP4 format may make deployment less demanding on memory than full-precision weights.

Those improvements do not mean the U.S. has caught the leaders in the broader open-weight field. The supplied Artificial Analysis comparison puts Super VL at 18th overall and 11th among permissively licensed models. The source says nine of the ten highest-scoring models in its table are from Chinese labs, while the UAE’s K2 Horizon 375B A23B scores 31. These standings are a snapshot of one evaluation index, not a measure of every model’s usefulness or deployment suitability.

Licensing also changes the comparison. Meta’s Muse Spark Open scored 39 in the same-day results, but the source describes its Muse Model License as having commercial restrictions. Buyers weighing models must compare not only benchmark scores, but also permitted uses, deployment options, price and performance on their own tasks.

How the Models Compare

The source describes Super VL as the flagship model and Bolt as the first in NVIDIA’s smaller Bolt family. Both are open-weight models, but their availability differs: Super VL is listed with an API price of $0.20 per million input tokens and $0.575 per million output tokens; Bolt is described as self-host only, with no API price listed at the time of the report. Artificial Analysis’s reported Super VL cost per index task was $0.25, while the Bolt figure was unavailable.

Super VL’s result also compares favorably with NVIDIA’s previous flagship, Nemotron 3 Ultra, which the source lists at 550 billion total parameters, 55 billion active parameters and an Intelligence Index score of 23. The newer model scored higher with fewer parameters and a lower reported task cost. Bolt, meanwhile, ranked fourth among models below 100 billion total parameters in the cited table and first among U.S. models in that size category.

The wider comparison includes models with different licenses. The source lists China’s MiMo-V2.6-Pro at 46 under an MIT license, several other Chinese models at 44 to 45 with commercial conditions, and the UAE’s K2 Horizon at 31 under Apache 2.0. Meta’s Muse Spark Open scored 39. These figures come from the source’s account of Artificial Analysis’s rankings on October 15, 2026, and may change as models or evaluations are updated.

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

— Thorsten Meyer, source report

Licenses, Costs and Rankings

The supplied source cuts off while describing the Muse Model License, so its precise commercial restrictions cannot be independently detailed here. The comparison also does not establish how the models perform across every real-world task, or whether benchmark gains will carry over to particular deployments. Bolt’s API availability and cost per index task were not reported, limiting a direct cost comparison with Super VL.

Rankings and serving measurements can change as providers update models and infrastructure. The source reports a pre-release DeepInfra test of Super VL at about 630 output tokens per second on 100,000-token prompts, but gives a median of 196 tokens per second across six listed providers. That gap underscores that reported speed depends on the service and test conditions. The source does not provide enough detail here to assess every evaluation’s methodology or the statistical uncertainty around scores.

Deployment Data Will Matter

The next useful evidence will come from broader availability and independent testing of the released weights. For Super VL, comparisons across hosting providers can clarify how the reported speed and cost translate into routine use. For Bolt, a public API or other published pricing would allow buyers to assess total cost rather than token efficiency alone.

Readers should also watch for updated Artificial Analysis listings and any fuller publication of licensing terms, particularly for Muse Spark Open. Until then, the confirmed picture is narrower: NVIDIA’s new models improved its benchmark position, while the supplied rankings still place higher-scoring Chinese and UAE models ahead overall.

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 in the comparison reported by the source.

Does Super VL lead all open-weight models?

No. The source places it 18th overall in the cited ranking and says several Chinese models and a UAE model scored higher. It describes Super VL as leading among U.S. permissively licensed models.

Can businesses use Super VL commercially?

The source says Super VL uses the OpenMDW-1.1 license, which allows commercial use. Users should review the license text for terms that apply to their specific use.

Is Bolt cheaper to run than Super VL?

The available source does not establish that. Bolt had no public API or measured cost per index task at the time of the report, while its output-token use per task was reported as higher than Nemotron 3 Ultra’s.

Source: ThorstenMeyerAI.com