What Nemotron 3.5’S Score Says About America’s Open-Weight Gap

🔍 Read the full analysis: What Nemotron 3.5’S Score Says About America’s Open-Weight Gap on ThorstenMeyerAI.com

Age 18–24?Offer from Amazon

Prime made for students and young adults

• Fast, free delivery for dorm and study essentials
• Prime Video and Amazon Music included
• Member-only deals

Try Prime for Young Adults →Free trial for eligible 18–24 year olds

As an affiliate, we earn on qualifying purchases.Hide

NVIDIA released two open-weight models on October 15, and an evaluation published the same day found that its Nemotron 3.5 Super VL scored 26.5 on Artificial Analysis’s Intelligence Index, roughly twice the 13 recorded for its predecessor. The result strengthens NVIDIA’s position among U.S. models with permissive licenses, but it does not close the gap with the highest-scoring open-weight models listed in the evaluation, most of which come from China.

Artificial Analysis reported a rounded Intelligence Index score of 26 for Super VL and 21 for Nemotron 3.5 Bolt 74B-A9B. Super VL has 120 billion total parameters, with 12 billion active for each token; Bolt has 74 billion total and 9 billion active. Both accept text and images and return text. NVIDIA describes Bolt as the first model in a family designed to run with less hardware.

The evaluation says Super VL used about 49,000 output tokens per index task, compared with about 114,000 for Nemotron 3 Super. Its measured task cost was $0.25, versus $1.64 for the predecessor. These are Artificial Analysis’s reported results for its evaluation and pricing assumptions, not a guarantee of costs or performance for every deployment. Super VL’s OpenMDW-1.1 license allows commercial use, according to the source. Bolt did not yet have a public API or a listed API price in the source material.

Artificial Analysis also tested Super VL in NVIDIA’s NVFP4 format. It reported that this version scored within 0.1 points of the full-precision BF16 weights on the overall index, with no individual evaluation changing by more than 2.6 points. For local operators, the finding suggests a way to reduce memory requirements while retaining similar measured performance; actual results may vary with hardware and implementation.

A Stronger U.S. Contender, Not a Leader

The result matters because it shows a substantial improvement in a model that can be used commercially under its stated license, while also highlighting how far the top of the evaluated field remains from it. Artificial Analysis placed Super VL 18th among the open-weight models it listed, and 11th when counting only permissively licensed models. The source’s central distinction is that “open-weight” does not always mean unrestricted commercial use.

The same-day comparison with Meta’s Muse Spark Open illustrates that distinction. Artificial Analysis gave Muse Spark Open a score of 39, above Super VL’s 26.5, but the source says Meta’s Muse Model License carries commercial restrictions. Muse Spark Open is also much larger, at 816 billion total parameters and 42 billion active. For companies choosing a model, the highest score may not be the only relevant measure: licensing terms, deployment costs, hardware needs and task-specific performance also affect whether a model is usable.

Super VL’s reported task cost and token reduction may make it attractive for some workloads, but those figures should not be treated as a universal price comparison. Bolt presents a separate trade-off: it has fewer active parameters than NVIDIA’s Nemotron 3 Ultra, but the evaluation found it generated about 82,000 output tokens per index task, compared with Ultra’s 28,000. Without a public API price for Bolt, its cost per completed task cannot yet be compared on the same basis.

How the Models Compare

The two new models are mixture-of-experts systems, meaning only part of the model is active for each token. Super VL has a one-million-token context window; Bolt’s is 256,000 tokens. Artificial Analysis reported Super VL’s API price as $0.20 per million input tokens and $0.575 per million output tokens, and calculated $0.25 per index task. Those figures reflect the provider and testing conditions available when the source was written.

Bolt’s score of 21 was 1.5 points below Nemotron 3 Ultra’s 23, despite Bolt having 74 billion total parameters compared with Ultra’s 550 billion. The source reports mixed results across individual evaluations: Bolt scored higher on AutomationBench-AA and AA-Briefcase, while Ultra scored higher on long-context reasoning and Humanity’s Last Exam. The comparison indicates that an aggregate index does not capture every model’s strengths.

The wider ranking in the source was led by Chinese models, including Xiaomi’s MiMo-V2.6-Pro at 46 and several models scoring between 44 and 45. The UAE’s K2 Horizon 375B A23B scored 31. Meta’s Muse Spark Open scored 39. The source says nine of the ten highest-scoring models came from Chinese labs, counting each model once; that ranking is specific to the Artificial Analysis list described in the October 16 report and may change as models or evaluations are updated.

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

— Thorsten Meyer, source article

Licenses, Prices and Ranking Limits

The source material ends partway through its discussion of the Muse Model License, so its full commercial restrictions are not specified here. It also does not provide a public API price for Bolt, preventing a direct comparison of its cost per task with Super VL or Nemotron 3 Ultra. The source’s rankings and costs describe Artificial Analysis’s evaluation and listed providers at the time; they do not establish how every model will perform across all tasks or deployment setups.

It is also unclear from the supplied material how the ranking handles every licensing category beyond the examples given, or whether all listed models were evaluated under identical serving conditions. Speed figures are particularly provider-dependent: the source cites a pre-release DeepInfra result of about 630 output tokens per second for Super VL with 100,000-token prompts, while the median across six listed providers was 196 tokens per second. Those figures refer to different serving observations, not a guaranteed speed for users.

Watch for Bolt Pricing and Updates

The next useful comparison will come when Bolt has public serving options and a price that can be tested against its long output-token use. Further provider measurements could also show how consistently Super VL’s speed and NVFP4 results hold across different hardware and deployments.

For now, readers and buyers can treat the October 15 evaluation as a dated snapshot: Super VL made a marked gain over its predecessor and leads the U.S. permissively licensed models discussed in the source, while the top global scores remain higher. Future model releases, updated benchmarks and clearer license terms could change that picture.

Key Questions

What score did Nemotron 3.5 Super VL receive?

Artificial Analysis reported a rounded score of 26 on its Intelligence Index, with an unrounded score of 26.5.

Did Super VL close the gap with the highest-scoring models?

No. The source reports higher scores for several models, including Chinese models and Meta’s Muse Spark Open. Super VL’s gain improved NVIDIA’s position, but did not make it the global leader in the listed ranking.

Can businesses use Super VL commercially?

The source says Super VL uses the OpenMDW-1.1 license, which allows commercial use. Users should check the license text and applicable terms before deployment.

How does Bolt compare with Super VL?

Bolt scored 21, compared with Super VL’s 26.5, and has 74 billion total parameters versus Super VL’s 120 billion. The source says Bolt had no public API or API price at the time, and reported that it used substantially more output tokens per index task than Nemotron 3 Ultra.

Source: ThorstenMeyerAI.com