A Closer Look At Nemotron 3.5 Bolt’s Score, Size, And Token Bill

🔍 Read the full analysis: A Closer Look At Nemotron 3.5 Bolt’s Score, Size, And Token Bill 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, an open-weight model that scored 21.4 on the Artificial Analysis Intelligence Index v4.3.2, 1.5 points below the company’s much larger Nemotron 3 Ultra 550B-A55B. Bolt has 74 billion total parameters and 9 billion active per token, compared with Ultra’s 550 billion total and 55 billion active, but it ranked fourth among models under 100 billion parameters and generated about 82,000 output tokens per index task.

The benchmark comparison comes from Artificial Analysis Intelligence Index v4.3.2, which evaluated Bolt’s reasoning version. The 21.4 score is close to Ultra’s 22.9, despite Bolt having roughly one-seventh of its total parameters and one-sixth of its active parameters. NVIDIA’s previous smaller model, Nemotron 3 Super, scored 12.8 on the index, according to the source report. Bolt is a mixture-of-experts model with a reported 256,000-token context window; it accepts text and images and produces text.

Bolt’s performance varies by task. Artificial Analysis results cited in the report put its AutomationBench-AA score at 15.5%, versus 3.0% for Ultra, and its AA-Briefcase Elo at 912 versus 876. Ultra scored higher on long-context reasoning, with 79% on AA-LCR against Bolt’s 70%, and on Humanity’s Last Exam, at 28% against 25%. Both models scored poorly on Terminal-Bench 4.0: Bolt recorded 0%, and Ultra 0.5%.

Artificial Analysis placed Bolt fourth among models under 100B parameters, and first among U.S. models in that category. The report lists Qwen3.8 27B at 33.7, K2 Horizon MoVA 36B-A4B at 25.3, and G9v3-39A5B at 21.8 ahead of Bolt. Its open weights are available for self-hosting. The source says Artificial Analysis listed no hosted API provider at the time, leaving no published API price, speed measurement or cost per task on that page.

A Smaller Model With a Longer Run

Bolt’s headline score suggests a smaller model can approach a larger predecessor on an aggregate benchmark, but the score does not show how quickly or efficiently it completes work. Its 82,000 output tokens per index task are nearly three times Ultra’s 28,000, according to the report. For self-hosted users, generated tokens consume inference time and hardware capacity; for hosted users, token volume can affect billing, although no provider price was available for Bolt in the material.

The task-level results also make the model’s potential use cases more specific. Its reported lead on AutomationBench-AA, which measures multi-step workflows across SaaS tools, may interest teams testing agentic workflows. But the reported 0% Terminal-Bench result does not support treating Bolt as a capable shell agent, and its lower long-context result points to a weakness on document-heavy tasks. A 256,000-token context limit describes how much text it can accept, not whether it can reliably reason over all of it.

For organizations tracking model origin, the report’s ranking of Bolt as the highest-scoring U.S. model under 100B parameters may also be relevant. That is a category-specific comparison, not a claim that Bolt leads all models of its size: it trails three entries in the reported table. Its image input is another distinguishing feature in that comparison, while performance and speed still require practical testing.

How Bolt Compares With NVIDIA Models

Bolt is the first model in NVIDIA’s new Nemotron Bolt family, released alongside Nemotron 3.5 Super VL on October 15, 2026. The source report describes Super VL as the release drawing more attention, while identifying Bolt as a notable engineering result because of its score relative to its parameter count. The comparison is not simply one of model size: Bolt’s better results on some agentic benchmarks coexist with weaker results on knowledge and long-document measures.

Artificial Analysis’s size-class table, as presented in the report, puts Qwen3.8 27B well ahead at 33.7 and K2 Horizon MoVA at 25.3. Bolt’s 21.4 is close to G9v3-39A5B’s 21.8, and above the other listed models in the excerpt. The ranking counts each model once at its best configuration, according to the source. NVIDIA’s earlier Nemotron 3 Super, released in March, scored 12.8 in the index, giving the report a company-line comparison as well as the Ultra comparison.

The report also notes that Bolt’s permissive license classification appears on the model page, but says the specific license name was not listed there. No hosted API provider or speed result was listed in the cited material. Those gaps matter when comparing deployment options: open weights permit self-hosting, but do not by themselves establish operating cost, throughput or suitability for a particular workload.

““For batch jobs that run overnight, that is fine. For anything a person waits for, it is not.””

— Thorsten Meyer, source report

Speed, Costs and Reliability Still Unknown

No measured inference speed or cost per task was available in the source material. 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 examples, not benchmark measurements of Bolt. Actual runtime depends on hardware, software and generation settings.

The reported index score also does not settle how Bolt will perform on specific deployments. The source describes its AA-Omniscience score as −19, a result related to correct answers and penalties for confident errors, but does not provide enough detail in the supplied excerpt to interpret the score fully. The specific license name, availability of hosted services, and independent replication of the benchmark results are also not established by the material. The benchmark figures should be treated as reported evaluation results, not proof of performance across all tasks.

Deployment Tests Will Settle the Trade-Off

The next useful evidence will be independent speed and task testing on the hardware organizations would actually use, along with published hosted-provider availability and pricing. Such results could show whether Bolt’s lower active parameter count offsets its higher token use in real workloads, or whether lengthy generation makes it slower or more expensive than alternatives.

Teams considering the model can compare it against their own tasks rather than rely on the aggregate index alone: agentic workflows, long-document analysis, image inputs and command-line work produce different benchmark profiles. NVIDIA or Artificial Analysis may publish further details, but the source material does not give a schedule for additional evaluations or service launches.

Key Questions

What is Nemotron 3.5 Bolt?

It is NVIDIA’s open-weight, 74-billion-parameter mixture-of-experts model, with 9 billion parameters active per token. The reported version accepts text and images, produces text and supports reasoning.

How close is Bolt to Nemotron 3 Ultra?

Bolt scored 21.4 on Artificial Analysis Intelligence Index v4.3.2, compared with Ultra’s 22.9. That is a 1.5-point gap on this index, not evidence that the models perform equally across every task.

Is Bolt the top model under 100 billion parameters?

No. The source report places it fourth in the under-100B category and first among U.S. models in that category. Three listed models scored higher overall.

Why does Bolt’s token count matter?

The report says Bolt generated about 82,000 output tokens per index task. More output can require more inference time on self-hosted hardware or contribute to API charges, but Bolt had no listed hosted provider or measured speed in the cited material.

What remains unknown about using Bolt?

The cited material does not establish real-world throughput, API pricing or cost per task. It also does not name the specific license, despite describing the license classification as permissive.

Source: ThorstenMeyerAI.com