Can Nemotron 3.5 Bolt Deliver Ultra-Class Scores In A Smaller Model?

🔍 Read the full analysis: Can Nemotron 3.5 Bolt Deliver Ultra-Class Scores In A Smaller Model? on ThorstenMeyerAI.com

Before you orderOffer from Amazon

Get hardware and tech essentials 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, and the open-weight model scored 21.4 on Artificial Analysis’ Intelligence Index v4.3.2, 1.5 points behind the company’s 550B-parameter Nemotron 3 Ultra. The result puts Bolt near NVIDIA’s previous flagship despite its smaller size, but it does not make Bolt a leader across its model class: Artificial Analysis ranks it fourth among models under 100B parameters, and its evaluation found unusually high output-token use.

Bolt is a mixture-of-experts model with 74 billion total parameters and 9 billion active per token. It accepts text and images, produces text, supports reasoning and has a 256,000-token context window. NVIDIA has made its weights available for self-hosting. Artificial Analysis classifies its license as permissive, though the model page does not name the specific license.

In the same Intelligence Index comparison, Nemotron 3 Ultra 550B-A55B scored 22.9, with 550 billion total and 55 billion active parameters. Bolt’s score is therefore 1.5 points lower, based on about one-seventh the total parameters and one-sixth the active parameters. NVIDIA’s Nemotron 3 Super, released in March with 120 billion total and 12 billion active parameters, scored 12.8 on the index.

The overall score masks differences across tasks. Bolt scored 15.5% on AutomationBench-AA, which tests multi-step workflows across software tools, compared with 3.0% for Ultra. But Ultra did better on long-context reasoning, scoring 79% on AA-LCR against Bolt’s 70%, and on several knowledge-oriented tests. Both models scored close to zero on Terminal-Bench 4.0, so the available results do not support treating Bolt as a capable shell agent.

Bolt’s Agent Strengths and Trade-offs

The result matters to developers evaluating whether a smaller model can handle some tasks associated with a much larger system. Bolt’s AutomationBench score suggests potential for multi-step tool workflows, while its image input and open weights give organizations options for self-hosting and applications that need visual input. Artificial Analysis ranks Bolt as the highest-scoring US model under 100B parameters in its comparison, a distinction that may matter to organizations tracking model origin or provenance.

That does not establish that Bolt is faster or cheaper to run than competing models. Artificial Analysis lists no hosted API provider for Bolt and has not published speed or cost-per-task measurements. Its 9B active parameter count may reduce computation per generated token compared with a dense model of similar total size, but that is an architectural consideration, not a measured performance result.

Output length is a practical constraint. Bolt generated about 82,000 output tokens per index task, according to the source’s reporting of Artificial Analysis data. That is nearly three times Ultra’s 28,000. For self-hosters, more tokens mean more generation time; for hosted services, they can affect usage costs. The source’s examples of roughly 14 minutes at 100 tokens per second and 27 minutes at 50 tokens per second are arithmetic illustrations, not measured Bolt speeds.

How Bolt Ranks Against Peers

Bolt is not the top model in Artificial Analysis’ under-100B comparison. The source lists Qwen3.8 27B at 33.7, K2 Horizon MoVA 36B-A4B at 25.3 and G9v3-39A5B at 21.8, ahead of Bolt at 21.4. The comparison counts each model once at its best configuration. Qwen3.8’s score is substantially higher, while K2 Horizon MoVA also scores above Bolt with fewer total and active parameters, according to the source.

Bolt’s notable distinction in that list is its position among US models, not a class-wide lead. It also accepts images, which the source says K2 Horizon MoVA and G9v3 do not. These differences can affect which model suits a particular use, but benchmark rankings alone do not establish performance on a developer’s own tasks.

Artificial Analysis evaluated Bolt’s reasoning version. The index score describes its performance on that evaluation, not every possible configuration or real-world workload. Its results also show uneven capabilities: a strong relative result on one agentic-workflow benchmark sits alongside weaker long-document performance and a negative score on AA-Omniscience, a test that rewards correct answers and penalizes confident errors.

Speed, Cost and Reliability Still Unknown

There is no published hosted-provider listing, speed measurement or cost per task for Bolt in the cited Artificial Analysis model page. The reported token count cannot show how long Bolt takes on a particular GPU, since actual throughput depends on hardware, software and configuration. The time figures above are estimates based on assumed generation rates, not test results.

The benchmark results also leave open how Bolt performs on users’ own tasks, how often it gives incorrect answers, and whether its high token use is typical across workloads. The source reports a score of −19 on AA-Omniscience, but the supplied material cuts off before giving further detail about that result. It does not provide enough information to quantify Bolt’s hallucination rate or general reliability.

Finally, the source does not identify the exact permissive license or provide a complete account of deployment requirements. Developers should check NVIDIA’s model materials for the applicable terms and technical details before adopting the weights.

What Deployment Tests Must Show

The next useful evidence would be independent speed and cost measurements on named hardware, alongside tests of output quality at shorter generation lengths. A hosted API listing, if one becomes available, could provide a clearer basis for comparing per-task cost and response time with other models.

For now, the reported benchmarks point to a model with a specific profile: relatively strong results on the tested tool-workflow benchmark, lower scores than Ultra on long-context reasoning and several knowledge tasks, and high token use during the index evaluation. Teams considering self-hosting can compare Bolt against their own representative prompts and hardware, while treating public benchmark scores as evidence about those tests rather than a guarantee of production performance.

Key Questions

What did Nemotron 3.5 Bolt score?

Bolt scored 21.4 on Artificial Analysis’ Intelligence Index v4.3.2. NVIDIA’s Nemotron 3 Ultra scored 22.9 on the same index.

Is Bolt the highest-ranked model under 100B parameters?

No. Artificial Analysis places it fourth among models under 100B in the comparison cited by the source. It is listed as the highest-scoring US model in that class.

What does 74B-A9B mean?

Bolt has 74 billion total parameters in a mixture-of-experts architecture, with 9 billion active per token, according to the source.

Is Bolt faster or cheaper to run than Nemotron Ultra?

That has not been established in the cited material. Artificial Analysis has not published a speed measurement or cost-per-task figure for Bolt. Its higher reported output-token count could mean more generation time or cost, depending on hardware and deployment.

Can Bolt be self-hosted?

The source says NVIDIA released Bolt’s open weights for self-hosting and that Artificial Analysis classifies its license as permissive. The exact license name is not listed in the cited model page.

Source: ThorstenMeyerAI.com