Why Nemotron 3.5’S Score Boost Doesn’t Settle America’s Open-Weight Gap

🔍 Read the full analysis: Why Nemotron 3.5’S Score Boost Doesn’t Settle America’s Open-Weight Gap on ThorstenMeyerAI.com

Buying for a business?Offer from Amazon

Get business pricing on monitors, keyboards and dev gear

• 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 Nemotron 3.5 Super VL and Nemotron 3.5 Bolt 74B-A9B on October 15, and same-day Artificial Analysis evaluations show a substantial improvement over the company’s previous models. Super VL scored 26 on the Intelligence Index, up from 13 for its predecessor, strengthening the U.S. position among permissively licensed models without closing the score gap with leading models from China and the UAE.

Super VL is a 120-billion-parameter mixture-of-experts model, with 12 billion parameters active for each token. Artificial Analysis reported an unrounded Intelligence Index score of 26.5. Its predecessor, Nemotron 3 Super, scored 13 in the same index. The evaluation also recorded about 49,000 output tokens per index task, compared with roughly 114,000 for the predecessor, and an estimated task cost of $0.25, down from $1.64.

NVIDIA’s second release, Bolt 74B-A9B, scored 21, with 74 billion total parameters and 9 billion active. Artificial Analysis placed it fourth among models with fewer than 100 billion total parameters and first among U.S. models in that size category. Bolt has no public API yet, so its cost per task has not been measured. The evaluation listed a 256,000-token context window; Super VL’s is one million tokens.

Artificial Analysis also tested Super VL in NVFP4, NVIDIA’s four-bit number format. The evaluator found that version within 0.1 points of the full-precision BF16 weights on the index, with no individual evaluation shifting by more than 2.6 points. Its reported performance varies by provider: a pre-release DeepInfra endpoint reached about 630 output tokens per second on 100,000-token prompts, while the median across six listed providers was 196.

U.S. Gains Leave a Score Gap

The release matters for buyers and developers seeking commercially usable open weights. Super VL uses the OpenMDW-1.1 license, which the source says allows commercial use, and its reported task cost is lower than those of NVIDIA’s earlier models. The NVFP4 results may also make local deployment more practical for operators constrained by GPU memory, though actual speed and cost depend on hardware and the serving provider.

The broader rankings qualify the U.S. gains. Artificial Analysis listed China’s MiMo-V2.6-Pro at 46 points and the UAE’s K2 Horizon 375B A23B at 31. Super VL scored 26. The source reports that nine of the ten highest-scoring open-weight models in the table came from Chinese labs, counting each model once. Those figures indicate that a stronger U.S. entry has not yet shifted the overall ranking described by the source.

Licenses affect how those comparisons apply to organizations. Meta’s Muse Spark Open scored 39 points on the same day, but uses the Muse Model License, which the source describes as carrying commercial restrictions. Super VL is therefore described as the leading U.S. model among permissively licensed options, not the top-scoring model across all open-weight releases. For organizations choosing a model, score, price, hardware needs and permitted uses all matter.

How the Models Compare

The October 15 releases are both reasoning models that accept text and images and return text. Both use mixture-of-experts designs, activating only part of the network for each token. NVIDIA positions Super VL as its flagship and Bolt as the first model in a smaller family intended to run on less hardware.

Artificial Analysis’s comparison puts Super VL above NVIDIA’s previous Nemotron 3 Ultra, which scored 23 despite having 550 billion total and 55 billion active parameters. Bolt scored 21, close to Ultra with far fewer parameters, but produced about 82,000 output tokens per index task, compared with Ultra’s 28,000. Bolt’s lower model size alone does not establish its cost per answer, and the lack of a public API leaves that price unknown.

The source says Super VL ranks 18th among all open-weight models in the Artificial Analysis listing and 11th among permissively licensed models. It also reports that the model edges Thinking Machines’ Inkling Small, at 25.7, among permissively licensed U.S. models. These standings refer to the rankings available at publication; model scores and listings can change as evaluations are added.

Limits of the Current Rankings

The cited rankings capture one evaluator’s index and model listing; the source does not establish that the scores predict performance across every workload. Provider speed figures also differ substantially, and the reported 630-token-per-second result came from a pre-release endpoint rather than the six-provider median.

Several practical comparisons remain incomplete. Bolt has no public API, and Artificial Analysis had not measured its cost per task. The source also says its license was not yet listed by Artificial Analysis. While the source characterizes Muse Spark Open’s license as commercially restrictive, it does not provide the full terms here. No evidence in the supplied material shows whether the release will change adoption, U.S. market share or the composition of future rankings.

Deployment and Rankings Ahead

Further evidence will come from public access to Bolt and provider pricing, alongside additional deployments of Super VL. Those would allow buyers to compare the models’ cost and speed under more consistent conditions. Updated evaluations could also change their positions as new models enter the listings.

For now, the confirmed development is a marked improvement in NVIDIA’s evaluated scores and task efficiency, especially for Super VL. Whether that progress translates into a narrower national gap will depend on future model releases, licensing terms and independent results. The available figures show a stronger U.S. option in the permissive category, while higher scores from China and the UAE remain in the cited table.

Key Questions

What did Nemotron 3.5 Super VL score?

Artificial Analysis reported a 26-point Intelligence Index score, or 26.5 before rounding, on its October 15, 2026 evaluation.

Did Super VL close the U.S. open-weight gap?

No. It improved NVIDIA’s standing and led the permissively licensed U.S. models cited, but the source lists China’s MiMo-V2.6-Pro at 46 points and a UAE model at 31.

Can businesses use Super VL commercially?

The source says Super VL uses the OpenMDW-1.1 license, which allows commercial use. Readers should consult the license itself for its applicable terms.

How does Bolt compare with Super VL?

Bolt scored 21 points, compared with Super VL’s 26, and has fewer total and active parameters. It has no public API yet, and its task cost was not measured.

Source: ThorstenMeyerAI.com