Nemotron 3.5 Bolt Scores Big In A Model One-Seventh The Size

🔍 Read the full analysis: Nemotron 3.5 Bolt Scores Big In A Model One-Seventh The Size on ThorstenMeyerAI.com

Before you orderOffer from Amazon

Get the little things that make your day 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, close to the 22.9 recorded by NVIDIA’s 550B Nemotron 3 Ultra. Bolt has 74 billion total parameters and activates 9 billion per token, compared with Ultra’s 550 billion total and 55 billion active parameters. The result is notable for the size difference, but Bolt ranked fourth among models below 100B parameters, and its evaluation used substantially more output tokens per task than Ultra’s.

Artificial Analysis’ index puts Bolt 1.5 points behind Nemotron 3 Ultra, a model with roughly seven times as many total parameters. The two models are not interchangeable: the benchmark results reported by ThorstenMeyerAI.com show Bolt ahead on some agent-oriented tests, while Ultra performs better on several knowledge and long-document evaluations. Bolt scored 15.5% on AutomationBench-AA, compared with Ultra’s 3.0%. On AA-LCR, a long-context reasoning test, Bolt scored 70% and Ultra 79%.

Bolt is a mixture-of-experts model with a 256,000-token context window. It accepts text and images and produces text. The reasoning version was evaluated, and NVIDIA has released open weights for self-hosting. Artificial Analysis lists no hosted API provider for Bolt, so the source reports no measured API price, speed or cost per task. The model page classifies its license as permissive, but does not name the license.

The benchmark results also show limits. Bolt scored 2.8% on GDP.pdf, below Ultra’s 5.0%, and 25% on Humanity’s Last Exam, compared with Ultra’s 28%. Both models scored close to zero on Terminal-Bench 4.0: Bolt recorded 0% and Ultra 0.5%. Those results do not support treating Bolt as a capable shell-based coding agent. The index score is an overall comparison, not evidence that Bolt matches Ultra across every task.

A Smaller Model for Agent Work

Bolt’s strongest reported result is its 15.5% AutomationBench-AA score. The benchmark tests multi-step workflows involving SaaS tools, and Bolt’s result was substantially higher than Ultra’s 3.0%. It also slightly exceeded the 14.8% reported for NVIDIA’s new Nemotron 3.5 Super VL. That makes Bolt a model to watch for organizations exploring agentic workflows, though the score alone does not show how well it will perform in every real-world deployment.

Its size and open weights may appeal to teams that want to run a model on their own infrastructure or document its provenance. The source describes Bolt as the highest-scoring US model under 100B parameters in the Artificial Analysis comparison. That is a narrower claim than being the best model in its size class: the listed leaders include Qwen3.8 27B at 33.7 and K2 Horizon MoVA 36B-A4B at 25.3, both above Bolt’s 21.4.

Parameter count does not determine practical efficiency on its own. Bolt activates fewer parameters per token than many larger models, which could reduce compute needs in some self-hosted settings. But Artificial Analysis has not published a speed measurement for Bolt, and the benchmark indicates that it generated about 82,000 output tokens per index task. That much longer generation can mean more waiting and resource use. The potential hardware advantage remains unverified until comparable speed and cost measurements are available.

How Bolt Compares With NVIDIA’s Line

Bolt is the first model in NVIDIA’s new Nemotron Bolt family. It arrived alongside Nemotron 3.5 Super VL, which drew more attention in the source report because it was described as the highest-scoring permissively licensed open-weight model from the US. Bolt’s appeal is different: it is a smaller model whose overall index result approaches that of NVIDIA’s previous flagship.

NVIDIA’s Nemotron 3 Super, released in March according to the source, has 120 billion total parameters and 12 billion active parameters. It scored 12.8 on the same index. Bolt’s 21.4 is 8.6 points higher, despite having 39% fewer total parameters. This comparison signals a sizable change within NVIDIA’s lineup, but it does not by itself establish lower operating costs or faster responses.

Across the under-100B comparison cited in the report, Bolt sits behind Qwen3.8 27B, K2 Horizon MoVA 36B-A4B and G9v3-39A5B, which scored 33.7, 25.3 and 21.8, respectively. Bolt’s position reflects the particular Artificial Analysis index and configurations described by the source; benchmark rankings may differ across tasks, model versions and evaluation methods.

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

— Thorsten Meyer, ThorstenMeyerAI.com

Speed, Cost and License Details

There is no published hosted-provider listing for Bolt in the source, leaving API pricing, measured speed and cost per task unavailable. The report’s estimates that 82,000 tokens would take about 14 minutes at 100 tokens per second or 27 minutes at 50 tokens per second are arithmetic illustrations, not measurements of Bolt’s actual throughput. Real generation speed depends on the hardware, software and serving setup.

The source also says the model page calls the license permissive without naming it. Prospective users therefore need to check the license terms directly before deployment. Benchmark results do not settle how Bolt will perform in production, how much hardware it needs, or whether its verbosity can be controlled without reducing task quality.

The source excerpt gives a negative AA-Omniscience score of −19 for Bolt and zero for Ultra, but does not provide enough explanation of individual errors or task-level outcomes to establish how often Bolt may give confident incorrect answers. The index result should not be read as a complete measure of hallucination risk.

Awaiting Real-World Speed Tests

The next useful evidence will be independent throughput and cost testing on specified hardware, alongside comparable evaluations of output quality and response time. A hosted API listing, if one becomes available, could provide another basis for comparing price and latency. Until then, users considering self-hosting will need to test the model on their own workloads rather than infer performance from parameter counts or index scores.

Teams evaluating Bolt for agents can start with the distinction in the reported results: it performed comparatively well on AutomationBench-AA but poorly on Terminal-Bench 4.0 and below Ultra on long-context reasoning. Further tests across their specific tools and documents will show whether those benchmark strengths match their needs. NVIDIA or evaluators may also publish more information on the license and deployment requirements; the source does not give a date for any such updates.

Key Questions

What did NVIDIA release?

NVIDIA released Nemotron 3.5 Bolt 74B-A9B on October 15, 2026. It is an open-weight mixture-of-experts model with 74 billion total parameters and 9 billion active per token.

How close did Bolt score to Nemotron 3 Ultra?

Bolt scored 21.4 on Artificial Analysis Intelligence Index v4.3.2, compared with 22.9 for Nemotron 3 Ultra. The index scores are close, but the models have different strengths across individual tests.

Is Bolt the top model under 100B parameters?

No. The source report places Bolt fourth among models under 100B parameters in the cited comparison. It describes Bolt as the highest-scoring US model in that class, while several listed models scored higher overall.

What is the main practical caveat?

Bolt generated about 82,000 output tokens per index task, nearly three times Ultra’s 28,000. That can affect waiting time and resource use, but the source has no measured Bolt speed or hosted API cost.

Can Bolt be used as a shell agent?

The reported Terminal-Bench 4.0 result was 0% for Bolt. That benchmark does not support relying on it for shell-agent work, though performance on other specific tasks requires separate testing.

Source: ThorstenMeyerAI.com