Nemotron 3.5 Doubles NVIDIA’s Score—Why The Gap Still Matters

🔍 Read the full analysis: Nemotron 3.5 Doubles NVIDIA’s Score—Why The Gap Still Matters 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, and an independent evaluation found that its Nemotron 3.5 Super VL scored 26 on the Intelligence Index, double the 13 recorded by its predecessor seven months earlier. The result strengthens NVIDIA’s position among US models with permissive licenses, but the top of the broader open-weight rankings remains led by models from China and the United Arab Emirates.

Artificial Analysis reported an unrounded score of 26.5 for Super VL, compared with 13 for Nemotron 3 Super. The models are in the same size class: both have 120 billion total parameters and 12 billion active parameters. Super VL used about 49,000 output tokens per index task, down from about 114,000 for its predecessor, and its reported cost per task fell from $1.64 to $0.25.

NVIDIA also released Nemotron 3.5 Bolt 74B-A9B, a smaller mixture-of-experts model with 74 billion total parameters and 9 billion active. It scored 21, which Artificial Analysis placed fourth among models under 100 billion total parameters and first among US models in that size group. Bolt produced about 82,000 output tokens per index task; it has no public API yet, so its API cost has not been measured.

Both models accept text and images and return text. Super VL has a reported one-million-token context window and an OpenMDW-1.1 license that allows commercial use. Artificial Analysis said its NVFP4 version scored within 0.1 points of full-precision BF16 weights overall, though individual evaluations differed by as much as 2.6 points. Bolt has a 256,000-token context window; Artificial Analysis had not listed a license for it in the supplied source material.

A Stronger US Option, Still Behind

For developers choosing models they can run or adapt, Super VL’s combination of score and reported task cost is a meaningful change. It scored above NVIDIA’s earlier Nemotron 3 Ultra, which has 550 billion total parameters and 55 billion active parameters, while Super VL has less than a quarter as many total parameters. The evaluation put Ultra at 23 points and $0.60 per index task.

The broader competitive picture is less favorable to the United States. Artificial Analysis ranked Super VL 18th among open-weight models and 11th when considering permissively licensed models. The source says nine of the ten highest-scoring models in its listed open-weight ranking came from Chinese labs; Meta’s Muse Spark Open, at 39, was the other model in that top ten. Muse Spark Open has commercial restrictions, while Super VL’s license allows commercial use. That distinction matters to buyers, but it does not erase the score gap.

Price and speed also depend on how a model is served. Artificial Analysis measured Super VL at about 630 output tokens per second on a pre-release DeepInfra endpoint using NVFP4 weights and long prompts. The median among six providers listed on the model page was 196 tokens per second. Those figures describe different service results, not a guaranteed speed for every deployment.

How the Rankings Compare

The October 15 release arrived alongside Meta’s Muse Spark Open, which Artificial Analysis scored at 39. It is much larger than Super VL, with 816 billion total parameters and 42 billion active parameters, and had a reported cost of $0.84 per index task. Its Muse Model License has commercial restrictions, according to the supplied source.

The source identifies China’s MiMo-V2.6-Pro, from Xiaomi, as scoring 46 under an MIT license, at $0.13 per index task. It also lists GLM-5.3, Step 5 and Kimi K3 at scores from 44 to 45, under licenses with commercial conditions. The UAE’s K2 Horizon 375B A23B scored 31 and is listed under Apache 2.0. These comparisons show why license terms and evaluation scores need to be read together.

Bolt’s score is close to Ultra’s 23, despite its much smaller parameter count. It scored higher than Ultra on AutomationBench-AA (15.5% versus 3.0%) and AA-Briefcase (912 versus 874 Elo), but lower on long-context reasoning (70% versus 79%) and Humanity’s Last Exam (25% versus 28%). The results vary by benchmark rather than showing a uniform advantage.

Costs and Licensing Still Vary

Bolt’s price and license remain unresolved in the supplied material: it has no public API, and Artificial Analysis had not listed a license. Its output-token count is also much higher than Ultra’s, so a low cost per token would not by itself establish a low cost per answer.

Reported speeds are provider-specific. The 630-token-per-second measurement came from a pre-release DeepInfra endpoint, while the listed-provider median was 196. The source does not establish that the faster result will be typical across providers or deployments. Rankings and prices may also change as more providers list the models and evaluations are updated.

Provider Listings Will Fill Gaps

The next practical milestones are public access to Bolt and publication of its license, along with provider pricing that would allow buyers to compare its task costs. Super VL’s speed and cost should also become easier to assess as more providers serve it under comparable conditions. Until then, the October 15 evaluation establishes a substantial gain for NVIDIA’s model line, while leaving the larger open-weight ranking gap in place.

Key Questions

What did Nemotron 3.5 Super VL score?

Artificial Analysis reported a score of 26 on its Intelligence Index, or 26.5 unrounded. The predecessor, Nemotron 3 Super, scored 13.

How much did Super VL cost per index task?

Artificial Analysis listed $0.25 per index task, compared with $1.64 for Nemotron 3 Super. The supplied source does not establish that this cost will apply across all providers or deployments.

Is Nemotron 3.5 Super VL commercially usable?

The source says Super VL uses the OpenMDW-1.1 license, which allows commercial use. Bolt’s license was not listed by Artificial Analysis in the supplied material.

Does NVIDIA lead the open-weight rankings?

No. Super VL leads the permissively licensed US models in the source’s comparison, but it ranked 18th among open-weight models. Higher-scoring models from China and the UAE remained ahead.

Source: ThorstenMeyerAI.com