A Closer Look At Nemotron 3.5 And America’s Open-Weight AI Gap

🔍 Read the full analysis: A Closer Look At Nemotron 3.5 And America’s Open-Weight AI Gap 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 models on October 15, 2026, as independent evaluations showed gains in performance and efficiency but left the company’s models behind the leading entries in the global open-weight field. Nemotron 3.5 Super VL scored 26 on Artificial Analysis’s Intelligence Index, while Nemotron 3.5 Bolt 74B-A9B scored 21.

Super VL is the larger model, with 120 billion total parameters and 12 billion active for each token. It accepts text and images, produces text, and has a reported context window of one million tokens. Its OpenMDW-1.1 license allows commercial use. Artificial Analysis listed API pricing at $0.20 per million input tokens and $0.575 per million output tokens, and estimated a cost of $0.25 per Intelligence Index task.

Bolt has 74 billion total parameters and 9 billion active, a 256,000-token context window, and the same text-and-image input and text output format. It is the first model in a family NVIDIA describes as designed to run with less hardware. Artificial Analysis had not listed a license or public API price for Bolt at publication, so its cost per task was not available.

Artificial Analysis reported that Super VL’s score was 26.5 before rounding, compared with 13 for the earlier Nemotron 3 Super. In the evaluation, Super VL used about 49,000 output tokens per task, versus about 114,000 for its predecessor; estimated task cost fell from $1.64 to $0.25. The evaluator also said an NVFP4 version scored within 0.1 points of the BF16 version overall, though results varied by individual evaluation.

Efficiency Gains, Ranking Limits

The release gives developers and organizations another option for running capable models with open weights and, for Super VL, a license that permits commercial use. Lower estimated task costs and reduced output-token use could matter to teams that run models repeatedly, while the NVFP4 result may make Super VL easier to host on systems with less memory than full-precision weights require.

The broader competitive picture is less conclusive. Artificial Analysis ranked Super VL 18th among the open-weight models it listed, and 11th when counting only permissively licensed models. The source article says the top of the overall list was led by Chinese models, with a UAE model also ahead of Super VL. These positions describe one evaluator’s index and model list at a particular time; they do not establish performance in every task or market.

Licensing also affects how readers interpret “open.” Super VL permits commercial use under OpenMDW-1.1, while some higher-ranked models have commercial conditions. Meta’s Muse Spark Open, released the same day, scored 39 but uses the Muse Model License, which the source describes as commercially restrictive. A ranking by score alone therefore does not answer which model is most usable for a particular organization.

From Nemotron 3 to 3.5

The source article describes the Super VL result as a substantial gain over Nemotron 3 Super, released seven months earlier. Both are in the 120-billion-total, 12-billion-active parameter class, according to the source, while the newer model scored 26 versus 13. Artificial Analysis’s reported output-token and cost figures also fell between the two evaluations.

NVIDIA’s earlier Nemotron 3 Ultra is a larger comparison point: it has 550 billion total and 55 billion active parameters, scored 23, and had an estimated task cost of $0.60 on the same index. Super VL scored higher in that evaluation despite its smaller parameter count. Bolt scored 21, close to Ultra’s score with far fewer parameters, though it used about 82,000 output tokens per task compared with Ultra’s 28,000.

The source also reports that Artificial Analysis measured about 630 output tokens per second for Super VL on a pre-release DeepInfra endpoint using NVFP4 weights and 100,000-token prompts. The median across six providers listed on the model page was 196 tokens per second. Those figures reflect particular serving setups, so actual speed can differ by provider and deployment.

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

— Thorsten Meyer, source article

Licenses, Costs and Real-World Results

Artificial Analysis’s index is one benchmark, and the supplied source does not establish how these models perform across all real-world uses. Rankings may change as the evaluator adds models or updates scores. The source does not provide the full evaluation methodology or enough detail to independently assess each comparison here.

Bolt’s public API availability, pricing and license were not listed in the source material. Its reported token use also leaves its likely cost per completed task unknown. Super VL’s speed varied across providers, and the source’s fastest measurement came from a pre-release endpoint, so it should not be treated as a universal serving rate.

The source’s global comparison depends on which models are counted as open-weight and on how commercial restrictions are treated. It reports that nine of the ten highest-scoring models came from Chinese labs, but the supplied material ends during its discussion of Meta’s Muse Model License. The full license terms and any further qualifications are therefore not available here.

API Access and Updated Rankings

For Super VL, the next practical measure is how its published performance and cost estimates hold up across public deployments and different workloads. Artificial Analysis’s model page lists results from multiple providers, and further updates could give buyers a broader view of serving speed and consistency.

For Bolt, a published license, API access and pricing would let developers compare its total operating cost with its benchmark score and assess whether it suits hosted or self-managed use. Until those details are available, its cost-per-task comparison remains incomplete.

Further model releases and benchmark updates will show whether the US position changes beyond NVIDIA’s own improved results. The current evaluation establishes a performance gain for these two models; it does not settle the wider competition.

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, according to the source.

How did the models score?

Artificial Analysis gave Super VL an Intelligence Index score of 26 (26.5 unrounded) and Bolt a score of 21 (21.4 unrounded). The source reports that Nemotron 3 Super scored 13.

Does Super VL lead all open-weight models?

No. The source places Super VL 18th in Artificial Analysis’s overall open-weight ranking and 11th among permissively licensed models. It describes Super VL as the leading US model among permissively licensed models, while higher-scoring models include entries from China and the UAE.

Can businesses use Super VL commercially?

The source says Super VL uses the OpenMDW-1.1 license, which allows commercial use. It says the Muse Spark Open license has commercial restrictions; readers should consult each model’s published license for the terms that apply to their use.

Is Bolt available through an API?

The source says no public API price was listed for Bolt at publication and describes it as self-host only. Its license and cost per benchmark task were also not listed in the supplied material.

Source: ThorstenMeyerAI.com