Nemotron 3.5’S Score Doubles, But America’s Open-Weight Lead Is Unclear

🔍 Read the full analysis: Nemotron 3.5’S Score Doubles, But America’s Open-Weight Lead Is Unclear on ThorstenMeyerAI.com

Buying for a business?Offer from Amazon

Get business pricing on tech for your team

• 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 reasoning models on October 15, and an evaluation published the same day by Artificial Analysis puts Nemotron 3.5 Super VL at 26 on its Intelligence Index—twice the score of its predecessor. The result strengthens NVIDIA’s position among US models with permissive licenses, while the wider open-weight rankings remain led by models from China and the UAE.

Nemotron 3.5 Super VL has 120 billion total parameters, with 12 billion active for each token, and accepts text and images as input. Artificial Analysis gives it an Intelligence Index score of 26 (26.5 before rounding), compared with 13 for the similarly sized Nemotron 3 Super. The evaluation reports about 49,000 output tokens per index task for Super VL, down from about 114,000 for its predecessor, and a cost of $0.25 per task, compared with $1.64.

The second release, Nemotron 3.5 Bolt 74B-A9B, has 74 billion total parameters and 9 billion active parameters. It scored 21. Artificial Analysis ranks it fourth among models below 100 billion total parameters and first among US models in that size group. Bolt has no public API listed in the source, so its cost per task has not been measured.

Both models accept text and images and return text. Super VL uses NVIDIA’s NVFP4 four-bit format. Artificial Analysis reported that its score was within 0.1 points of the model’s full-precision BF16 version, with no individual evaluation differing by more than 2.6 points. The source lists Super VL under the OpenMDW-1.1 license, which allows commercial use; Artificial Analysis had not listed a license for Bolt.

A Stronger US Option, With Limits

Super VL’s score and reported task cost give buyers and developers a more capable US open-weight option in the efficient middle of the market. Its listed API price is $0.20 per million input tokens and $0.575 per million output tokens. The source also reports an especially fast result on a pre-release DeepInfra endpoint serving NVFP4 weights: about 630 output tokens per second with 100,000-token prompts and 1.8 seconds to the first token. That is a provider-specific measurement; the median speed among six providers listed on the model page was 196 tokens per second.

The release does not put the US at the top of the broader rankings cited in the source. Super VL is 18th among the listed open-weight models and 11th among permissively licensed models. The distinction matters because license terms affect whether companies can use a model commercially, while benchmark scores alone do not establish which model will work best for a particular task. Super VL leads the permissively licensed US models in this comparison, but that lead does not extend to the global ranking.

Bolt offers another measure of efficiency: its score of 21 is close to NVIDIA’s Nemotron 3 Ultra score of 23 despite Bolt having far fewer parameters. Its results vary by evaluation, however. Bolt scored higher than Ultra on AutomationBench-AA and AA-Briefcase, while trailing on long-context reasoning and Humanity’s Last Exam. It also produced about 82,000 output tokens per index task, compared with Ultra’s 28,000. Without a public API price, the cost of those longer answers is not yet established.

How the Rankings Compare

The Artificial Analysis Intelligence Index is the basis for the scores cited here; those figures are evaluation results, not a direct measure of performance on every real-world task. In the rankings described by the source, China’s MiMo-V2.6-Pro scored 46, while GLM-5.3, Step 5 and Kimi K3 scored between 44 and 45. The source says nine of the ten highest-scoring listed open-weight models came from Chinese labs. The UAE’s K2 Horizon 375B A23B scored 31.

Meta released Muse Spark Open on the same day as NVIDIA’s models. It scored 39, with 816 billion total parameters and 42 billion active parameters, and a reported cost of $0.84 per index task. Its license is described in the source as the Muse Model License, with commercial restrictions. That makes license terms a key part of comparisons: a higher score does not by itself mean a model is available on the same terms as Super VL.

NVIDIA’s own predecessor comparison is more direct: Nemotron 3 Super scored 13 seven months earlier, at the same 120-billion-total and 12-billion-active parameter scale. The supplied figures show a doubled rounded score alongside fewer output tokens and lower reported task cost. These are Artificial Analysis measurements and price estimates cited in the source, rather than a guarantee of results or costs across all deployments.

License and Deployment Questions

The source does not provide a listed license or public API price for Bolt, so commercial-use terms and comparable hosted costs remain unknown. Super VL’s speed result came from a pre-release DeepInfra endpoint; performance on other providers may differ, as the reported provider median was substantially lower.

The rankings are tied to Artificial Analysis’s listed models and its Intelligence Index. The supplied material does not explain how all license categories are handled in every ranking, or whether the rankings have changed since publication. Nor does a single index score establish performance for a specific workload. The Muse Spark Open license is described as carrying commercial restrictions, but the source excerpt ends before giving its full terms.

More Pricing and Ranking Data

Further information that would clarify the practical comparison includes a public API and price for Bolt, a listed license, and performance measurements across additional providers. Buyers comparing Super VL with Muse Spark Open or higher-scoring models will also need to check the applicable license terms and test models on their own tasks.

The rankings may change as Artificial Analysis evaluates more models or updates existing results. For now, the published figures show a substantial improvement over NVIDIA’s prior model and a stronger position among permissively licensed US options, while the highest scores in the cited open-weight table remain elsewhere.

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 are open-weight reasoning models that 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 before rounding. The source says Nemotron 3 Super scored 13 seven months earlier at the same stated parameter scale.

Does Super VL lead all open-weight models?

No. The source places it 18th among the listed open-weight models and 11th among permissively licensed models. It describes Super VL as the leading US model among permissively licensed options.

Can companies use Super VL commercially?

The source lists Super VL under OpenMDW-1.1 and says commercial use is allowed. It does not list a license for Bolt.

How much does Bolt cost to run?

The source says Bolt had no public API and Artificial Analysis had not measured a cost per index task. Its cost cannot be compared with Super VL’s reported $0.25 per task from the supplied figures.

Source: ThorstenMeyerAI.com