Nemotron 3.5 Makes A Score Leap, But America’s Open-Weight Gap Persists

🔍 Read the full analysis: Nemotron 3.5 Makes A Score Leap, But America’s Open-Weight Gap Persists 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 them at 26 and 21 on its Intelligence Index. Super VL doubled the score of its predecessor, but the rankings still put leading Chinese models and one UAE model above it, leaving the US behind in the broader open-weight field.

Artificial Analysis reports that Super VL scored 26.5 before rounding, compared with 13 for Nemotron 3 Super, which was evaluated seven months earlier. Both models are mixture-of-experts reasoning systems that accept text and images and return text. Super VL has 120 billion total parameters, with 12 billion active per token, and a one-million-token context window. Bolt has 74 billion total parameters, 9 billion active, and a 256,000-token window.

Artificial Analysis estimates Super VL uses about 49,000 output tokens per index task, down from about 114,000 for its predecessor. Its estimated cost per task fell from $1.64 to $0.25. The organization also measured the NVFP4 version within 0.1 points of the full-precision BF16 version on the overall index; no individual evaluation differed by more than 2.6 points. NVIDIA says NVFP4 uses a four-bit number format, which can reduce memory requirements compared with 16-bit weights.

Bolt scored 21, 1.5 points below NVIDIA’s larger Nemotron 3 Ultra, which scored 23. Artificial Analysis ranks Bolt fourth among models with fewer than 100 billion total parameters and first among US models in that size group. Its output per index task was about 82,000 tokens, compared with 28,000 for Ultra. Artificial Analysis had no public API price for Bolt at publication, so its cost per task was not measured.

US Models Gain Ground in Efficiency

The results make a case for NVIDIA’s models in settings where cost, hardware use and permissive licensing matter. Artificial Analysis lists Super VL at $0.20 per million input tokens and $0.575 per million output tokens, with an estimated $0.25 cost per index task. The assessment suggests that a smaller, more economical model can improve substantially without matching the highest scores in the global rankings.

The ranking gap remains substantial. Artificial Analysis lists Xiaomi’s MiMo-V2.6-Pro at 46 and an open-weight model from the UAE’s Institute of Foundation Models at 31. Super VL’s score is 20 points below MiMo-V2.6-Pro. Its position is stronger when the comparison is limited to permissively licensed models: Artificial Analysis calls it the leading US model in that category. That narrower result does not make it the overall leader.

License terms shape who can use a model and how. NVIDIA’s OpenMDW-1.1 license allows commercial use, according to the source material. The source says Meta’s Muse Spark Open scored 39 but uses a license with commercial restrictions. For organizations considering adoption, benchmark scores and inference costs are only part of the decision; the license can affect whether a model fits a product or service.

How the Rankings Compare

The October 15 evaluations place the NVIDIA releases within a wider set of open-weight models. Artificial Analysis lists several Chinese models between 44 and 46, including models with commercial conditions in their licenses, and the UAE’s K2 Horizon 375B A23B at 31 under Apache 2.0. The source material says nine of the ten highest-scoring models in the listing are from Chinese labs; Meta’s Muse Spark Open, at 39, is the other model in that group.

Super VL ranks 18th among all open-weight models in the table described by the source, and 11th when limited to permissively licensed models. Those positions depend on the models and scores included in the Artificial Analysis listing at the time. Within NVIDIA’s own line, Super VL also scored above Nemotron 3 Ultra, which has 550 billion total and 55 billion active parameters, and had an estimated cost per index task of $0.60.

For speed, Artificial Analysis measured a pre-release DeepInfra endpoint serving NVFP4 weights at about 630 output tokens per second with 100,000-token prompts, and 1.8 seconds to first token. The source describes that as the fastest measured result at that prompt length among models scoring 20 or higher. Across six providers listed on the model page, the median speed was 196 tokens per second, indicating that reported speed depends on the provider and serving setup.

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

— Thorsten Meyer, ThorstenMeyerAI.com

Prices and Licenses Still Vary

Bolt had no public API listed in the source material, leaving its commercial inference cost unknown. Its long output length also means the price per token alone would not establish its cost per answer. The reported speed measurement came from one pre-release DeepInfra endpoint; the source’s six-provider median is lower, and results may vary by provider.

The source says Muse Spark Open uses a license with commercial restrictions but does not provide the license terms in full. It also says Artificial Analysis had not listed a license for Bolt. The rankings reflect the listed models and evaluations at publication; how they may change as more models are assessed is not established in the source.

Provider Pricing and Rankings Ahead

The next practical markers are whether Bolt gains a public API and published pricing, and whether more providers list Super VL with comparable performance data. Those details will help buyers judge the cost of using the models beyond benchmark tasks. Further model evaluations may also shift the open-weight rankings, but the source material gives no timetable for upcoming releases or updates.

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 produce text.

How did Super VL score against its predecessor?

Artificial Analysis scored Super VL at 26 on its Intelligence Index, or 26.5 before rounding. Nemotron 3 Super scored 13.

Does Super VL lead all open-weight models?

No. The source’s Artificial Analysis ranking places Super VL 18th among listed open-weight models. It describes Super VL as the leading US model among permissively licensed models.

How much does Bolt cost to use through an API?

The source material says Bolt had no public API and no measured API price at publication. Its cost per index task was therefore not available.

Source: ThorstenMeyerAI.com