Can Nemotron 3.5’S Score Narrow America’s Open-Weight Gap?

Read the full analysis: Can Nemotron 3.5’S Score Narrow America’s Open-Weight Gap? on ThorstenMeyerAI.com

NVIDIA released Nemotron 3.5 Super VL and Nemotron 3.5 Bolt on October 15, with same-day testing by Artificial Analysis putting Super VL at 26 on its Intelligence Index. That is twice the score of its predecessor, but the result does not put the United States at the top of the global open-weight rankings: leading models from China and the UAE score higher.

Artificial Analysis gave Super VL an unrounded score of 26.5, compared with 13 for Nemotron 3 Super. The new model remains in the same reported size class, with 120 billion total parameters and 12 billion active. The evaluator also measured about 49,000 output tokens per index task, down from about 114,000 for the predecessor, and a task cost of $0.25 versus $1.64. These are results from Artificial Analysis’s evaluation and cost methodology, not a guarantee of performance or price across every deployment.

The second release, Nemotron 3.5 Bolt 74B-A9B, scored 21. It has 74 billion total parameters and 9 billion active parameters, compared with 550 billion and 55 billion respectively for NVIDIA’s Nemotron 3 Ultra, which scored 23. Artificial Analysis ranked Bolt fourth among models with fewer than 100 billion total parameters and first among US models in that group. It performed better than Ultra on two listed tests and worse on two others, including long-context reasoning and Humanity’s Last Exam.

Both models accept text and images and generate text, and use a mixture-of-experts design. Super VL is available under NVIDIA’s OpenMDW-1.1 license, which the source says allows commercial use. Its NVFP4 version scored within 0.1 points of full-precision BF16 weights in Artificial Analysis’s testing. Bolt did not yet have a public API listed in the supplied material, leaving its hosted price unmeasured.

A Better US Model, Not a Global Lead

The release offers a stronger option for developers seeking an open-weight US model, particularly where commercial licensing and deployment cost matter. Super VL’s reported task cost is substantially below its predecessor’s in the same evaluation, while its score is higher. Its NVFP4 result may also interest operators constrained by GPU memory, though actual performance depends on hardware and implementation.

The international comparison limits what the score can establish. Artificial Analysis lists China’s MiMo-V2.6-Pro at 46 and the UAE’s K2 Horizon 375B A23B at 31. The supplied report says Super VL ranks 18th among listed open-weight models, or 11th when counting only permissively licensed models. It describes Super VL as the leading US model among permissively licensed competitors, not the leading open-weight model overall. The distinction matters because licensing determines who can use a model commercially, while benchmark scores measure performance on a particular evaluation.

How the Scores Compare

The release follows NVIDIA’s Nemotron 3 Super, which Artificial Analysis scored at 13. The new Super VL’s reported improvement is therefore a comparison against a specific predecessor on the same index, rather than evidence that every task or deployment will see a twofold gain. The index combines multiple evaluations; individual test outcomes can differ, as the comparison between Bolt and Nemotron 3 Ultra illustrates.

The source also reports a speed result from a pre-release DeepInfra endpoint serving NVFP4 weights: about 630 output tokens per second with 100,000-token prompts and 1.8 seconds to first token. It says this was the fastest result measured at that prompt length among models scoring at least 20. The median speed across six providers listed on the model page was 196 tokens per second, underscoring that serving conditions affect results. Meta released Muse Spark Open on the same day, scoring 39 in the cited evaluation, but under a license described as carrying commercial restrictions.

Deployment Costs and License Details

Real-world costs and speeds remain provider-dependent. The reported $0.25 per index task is an evaluation estimate, and the cited speed test used a pre-release endpoint; the supplied information gives a lower median across six listed providers. Bolt has no public API price in the material, so its cost per answer cannot yet be compared with Super VL or Ultra. Its long output—about 82,000 tokens per index task, according to the report—could affect the cost of serving it.

The source material is incomplete on Muse Spark Open’s license terms: it identifies the Muse Model License and says it has commercial restrictions, but the supplied text ends before detailing them. The exact conditions, and how users should compare that license with OpenMDW-1.1, cannot be established from the information provided. Rankings also reflect the models and versions listed by Artificial Analysis at the time of testing and may change as new evaluations are added.

Watch for Public Serving Data

The next practical milestones are broader availability and published serving information, particularly for Bolt’s API pricing and independent results from additional providers. Those details would help users judge whether its smaller active parameter count translates into lower costs for their workloads, given its high output-token use in the cited index task.

Readers should also watch for updates to Artificial Analysis’s rankings and further licensing detail for competing releases. Until then, the supported conclusion is limited: NVIDIA made a substantial measured gain with Super VL and added a smaller Bolt model, while the cited global rankings still place higher-scoring open-weight systems outside the United States.

Key Questions

What did Nemotron 3.5 Super VL score?

Artificial Analysis scored it 26, with an unrounded result of 26.5, on its Intelligence Index. Its predecessor, Nemotron 3 Super, scored 13 in the same reported comparison.

Does Super VL lead the global open-weight rankings?

No. The supplied rankings place models from China and the UAE above it. The report describes Super VL as the leading US model among permissively licensed models, not the global leader.

How does Bolt differ from Super VL?

Bolt scored 21 and has 74 billion total parameters, including 9 billion active parameters. Super VL scored 26 and has 120 billion total parameters, including 12 billion active parameters. Bolt had no public API price listed in the source material.

Can companies use Super VL commercially?

The supplied source says Super VL uses NVIDIA’s OpenMDW-1.1 license, which allows commercial use. Users should review the license text itself to confirm how its terms apply to a specific use.

Source: ThorstenMeyerAI.com