Nemotron 3.5 Bolt Balances Compact Size, Strong Scores, And Token Use

🔍 Read the full analysis: Nemotron 3.5 Bolt Balances Compact Size, Strong Scores, And Token Use on ThorstenMeyerAI.com

Before you orderOffer from Amazon

Get the latest gadgets delivered free with Prime

• Fast, free delivery on millions of items
• Prime Video, Amazon Music and more included
• Member-only deals all year

Start your free Prime trial →Free trial for eligible customers · Cancel anytime

As an affiliate, we earn on qualifying purchases.Hide

NVIDIA released Nemotron 3.5 Bolt 74B-A9B on October 15, presenting a smaller open-weight model that scored 21.4 on Artificial Analysis Intelligence Index v4.3.2, close to the 22.9 score of its much larger Nemotron 3 Ultra. The result comes with trade-offs: Bolt ranks fourth among models under 100 billion parameters and used about 82,000 output tokens per index task, nearly three times Ultra’s total.

Bolt uses a mixture-of-experts architecture with 74 billion total parameters and 9 billion active for each token. It accepts text and images, produces text, supports reasoning, and has a 262,144-token context window. NVIDIA has released open weights for self-hosting. Artificial Analysis evaluated the reasoning version; its model listing showed no hosted API provider, speed measurement, price, or cost per task at the time described in the source.

The comparison with Nemotron 3 Ultra is close on the overall index but not across every task. Bolt scored 15.5% on AutomationBench-AA, which tests multi-step work across software tools, compared with Ultra’s 3.0%. It also scored higher on AA-Briefcase and CritPt in the listed results. Ultra led on long-document reasoning, with 79% on AA-LCR against Bolt’s 70%, and on several knowledge-related evaluations, including GDP.pdf and Humanity’s Last Exam. Both models scored poorly on Terminal-Bench 4.0, according to the figures provided.

Artificial Analysis ranked Bolt first among US models in the under-100B class, but fourth overall. The listed leaders were Qwen3.8 27B at 33.7, K2 Horizon MoVA 36B-A4B at 25.3 and G9v3-39A5B at 21.8; Bolt scored 21.4. Those comparisons show that a near-flagship score within NVIDIA’s own lineup does not make Bolt the leading model in its broader size category.

A Smaller Model With Measurable Trade-offs

Bolt’s results may appeal to organizations looking for an open-weight US model that can be run on their own infrastructure and can process images. Its 9 billion active parameters also offer a potential compute advantage over larger models, although that is an architectural implication, not a measured speed result. Artificial Analysis had not published Bolt’s inference speed, so users cannot yet compare real-world response times from the supplied evaluation details.

The token count complicates the efficiency case. Artificial Analysis reported that Bolt generated about 82,000 output tokens per index task, compared with 28,000 for Nemotron 3 Ultra. At hypothetical generation rates of 100 tokens per second, 82,000 tokens would take about 14 minutes; at 50 tokens per second, about 27 minutes. These are arithmetic estimates, not measured Bolt runtimes. The difference matters for interactive use, where users wait for answers, and for self-hosting, where long outputs consume GPU time.

For agentic workflows, the AutomationBench result is a stronger point in Bolt’s favor than the overall index alone. But the benchmark does not establish that Bolt will perform well in every tool environment, and the weak Terminal-Bench result argues against treating it as a general-purpose shell agent. The findings point to a model with a distinct task profile rather than a uniformly smaller substitute for Ultra.

How Bolt Fits NVIDIA’s Model Line

NVIDIA released two open-weight models on October 15: Nemotron 3.5 Super VL and Bolt, the first model in a new Bolt family. The source describes Super VL as the highest-scoring permissively licensed open-weight model from the US, but most of its detailed comparison focuses on Bolt. Artificial Analysis classified Bolt’s license as permissive, while the license name was not listed on the model page.

Bolt’s overall score is 1.5 points below Nemotron 3 Ultra, which has 550 billion total and 55 billion active parameters. The source describes Bolt as having roughly one-seventh of Ultra’s total parameters and one-sixth of its active parameters. Compared with Nemotron 3 Super, a 120B-total, 12B-active model released in March, Bolt scored 8.6 points higher while using 39% fewer total parameters, according to the source’s comparison.

Its fourth-place ranking also puts the headline score in perspective. Qwen3.8 27B scored 33.7, and K2 Horizon MoVA scored 25.3, both ahead of Bolt. The source notes that some higher-ranked competitors do not accept images, while Bolt does. That makes task requirements and deployment constraints relevant alongside the index ranking.

““For a 9B-active model, that is a real result.””

— Thorsten Meyer, ThorstenMeyerAI.com

Speed, Pricing and Licensing Details

Important deployment measures were not available in the source: measured inference speed, a hosted API price, and cost per task. The estimated generation times based on 50 or 100 tokens per second are calculations, not benchmarks of Bolt running on specific hardware. Actual performance will depend on the system, software and settings used.

The exact license name was also not listed on the model page, despite its permissive classification. The supplied material does not specify independent replication of the evaluation results or detail how Bolt performs across different deployment configurations. The source’s comparison tables describe benchmark outcomes, not guaranteed results in production.

The source material ends during its discussion of Bolt’s negative AA-Omniscience score, reporting a value of −19 but not providing the remainder of that explanation. The score is presented as a weakness in knowledge and resistance to confident wrong answers, but the available material does not give enough detail to characterize its full interpretation or the test methodology.

Deployment Tests Will Clarify Practical Use

The next useful evidence will come from independent speed and runtime measurements on specified hardware, along with testing of output length and quality across practical tasks. Those results would help determine whether Bolt’s active-parameter count translates into faster or less costly use despite its high token consumption.

Readers considering self-hosting should also check the published model files and license terms directly before deployment, and weigh the task-specific scores against their own workflows. Until hosted availability and pricing are listed, the source does not support a cost comparison with API models. Further benchmark results or deployment reports may clarify whether Bolt’s agentic performance holds outside the reported evaluation.

Key Questions

What is Nemotron 3.5 Bolt?

It is NVIDIA’s open-weight model released on October 15, 2026. It has 74 billion total parameters, 9 billion active per token, accepts text and images, and produces text.

How did Bolt compare with Nemotron 3 Ultra?

Bolt scored 21.4 on Artificial Analysis Intelligence Index v4.3.2, compared with Ultra’s 22.9. Bolt did better on AutomationBench-AA, while Ultra led on several knowledge and long-context evaluations listed in the source.

Is Bolt the top model under 100 billion parameters?

No. Artificial Analysis ranked it fourth in that class and first among US models. The listed scores for Qwen3.8 27B, K2 Horizon MoVA and G9v3-39A5B were all higher.

Why might Bolt take a long time to answer?

The source reports about 82,000 output tokens per index task, nearly three times Nemotron 3 Ultra’s 28,000. Estimated completion times based on assumed token rates are not measured Bolt runtimes; actual speed remains unreported in the cited material.

Can Bolt be used through a hosted API?

The source says Artificial Analysis listed no hosted API provider at the time. It reports open weights for self-hosting, but does not provide an API price or a measured cost per task.

Source: ThorstenMeyerAI.com