What To Know About Nemotron 3.5 Bolt’s Size And Token Costs

🔍 Read the full analysis: What To Know About Nemotron 3.5 Bolt’s Size And Token Costs on ThorstenMeyerAI.com

Before you orderOffer from Amazon

Get smart everyday buys 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 Oct. 15, an open-weight model that scored 21.4 on the Artificial Analysis Intelligence Index v4.3.2, close to the 22.9 score of the company’s much larger Nemotron 3 Ultra 550B-A55B. The comparison comes with important qualifications: Bolt ranks fourth among models under 100 billion total parameters and generated about 82,000 output tokens per index task, while no hosted API price or speed measurement is listed.

Bolt is a mixture-of-experts model with 74 billion total parameters and 9 billion active for each token, according to the source report. It has a 256,000-token context window, accepts text and images, produces text, and includes a reasoning version evaluated by Artificial Analysis. Its weights are available for self-hosting. The model page reportedly classifies its license as permissive but does not name the license.

Artificial Analysis’ index score puts Bolt 1.5 points below Nemotron 3 Ultra, which has 550 billion total and 55 billion active parameters. Bolt’s score is also 8.6 points above Nemotron 3 Super, the 120B-total, 12B-active model NVIDIA released in March, according to the source. These index results summarize a set of evaluations; they do not mean Bolt performs equally well across every task.

The reported task results show a mixed profile. Bolt scored 15.5% on AutomationBench-AA, compared with 3.0% for Ultra, and recorded an AA-Briefcase Elo score of 912 versus 876. Ultra did better on long-context reasoning: Bolt scored 70% on AA-LCR, compared with Ultra’s 79%. Bolt also scored lower on Humanity’s Last Exam, 25% to 28%, and GDP.pdf, 2.8% to 5.0%. Both models scored near zero on Terminal-Bench 4.0, so the available results do not support treating Bolt as a capable shell agent.

Token Volume Shapes Bolt’s Practical Cost

The headline score makes Bolt relevant to teams comparing model quality with self-hosting demands. Its 9 billion active parameters suggest a lower per-token compute requirement than a dense model of similar total size, but that architectural point is not a measured speed result. Artificial Analysis lists no speed measurement for Bolt, and there is no hosted provider listed on the model page, leaving API pricing and measured serving performance unavailable.

The token count adds another part to the cost calculation. Artificial Analysis reports about 82,000 output tokens per index task, nearly three times Nemotron 3 Ultra’s 28,000. The source gives illustrative arithmetic, not benchmarked latency: at 100 tokens per second, 82,000 tokens would take about 14 minutes to generate; at 50 tokens per second, about 27 minutes. Actual generation time depends on hardware, serving setup, and workload.

That distinction matters for users. A long-running batch job may tolerate more generated tokens if the result is useful, while an interactive application may be constrained by the time a person waits. Bolt’s score is therefore best read alongside its output volume, task-specific results, and deployment conditions, not as a direct measure of quality per minute or cost per answer.

Bolt’s Place Among Smaller Models

The source report places Bolt fourth among models under 100 billion total parameters on the Artificial Analysis index, and first among US models in that size category. The listed leaders are 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. Those rankings and scores reflect the model configurations and index information cited in the report, rather than a guarantee of performance on every use case.

Bolt’s result is notable within NVIDIA’s own model line: it approaches Ultra’s index score with a fraction of its total and active parameters. But the comparisons are not interchangeable. Bolt reportedly performs much better on AutomationBench-AA, while Ultra scores higher on several knowledge and long-document measures. Its 256,000-token context limit describes how much input it can accept; the lower AA-LCR score shows that context capacity alone does not establish how well it reasons over long material.

The source identifies image input as another distinction: Bolt accepts images as well as text, while two of the higher-ranked models it compares against do not. It also says Bolt’s 9B active-parameter design may reduce computation per generated token relative to a 27B dense model. Because Artificial Analysis has not measured Bolt’s speed, that remains an architectural expectation, not a confirmed throughput advantage.

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

— Thorsten Meyer, ThorstenMeyerAI.com

Pricing, Speed and License Details

Real-world latency and cost are not established in the available material. Artificial Analysis lists no hosted API provider, price, or speed measurement for Bolt. The reported 14- and 27-minute generation times are arithmetic examples based on assumed rates of 100 and 50 tokens per second, not tests of Bolt on specified hardware.

The source does not identify the name of Bolt’s permissive license, although it says the model page classifies the license that way. It also does not provide enough detail here to assess how the model’s benchmark results change across prompting strategies, hardware, or deployment settings. The reported index score and per-task token count should not be treated as a complete account of those conditions.

Some task-level outcomes are also weak: the source reports an AA-Omniscience score of −19 and says Terminal-Bench 4.0 results for Bolt and Ultra are both effectively zero. These figures indicate limitations in the cited evaluations, but they do not by themselves settle performance on other knowledge or software tasks. Use-case-specific testing remains necessary.

Watch for Independent Serving Tests

The next useful evidence would be measured inference speed and resource use on specified GPUs, including how output length affects latency for interactive and batch workloads. A hosted provider listing with published prices would allow readers to compare per-task costs directly; neither is included in the source material as of Oct. 16, 2026.

Readers assessing a self-hosted deployment can compare Bolt with alternatives using their own prompts and hardware, recording output tokens, completion quality, and elapsed time together. Further clarification of the named license would also help organizations review whether the model fits their intended use. Until those details are available, the confirmed picture is a near-flagship index score with a high reported token count, alongside specific strengths and weaknesses across evaluation tasks.

Key Questions

How large is Nemotron 3.5 Bolt?

Bolt has 74 billion total parameters, with 9 billion active per token, and is built as a mixture-of-experts model, according to the source report.

How does Bolt’s index score compare with Nemotron 3 Ultra?

Artificial Analysis scored Bolt 21.4 and Nemotron 3 Ultra 22.9 on Intelligence Index v4.3.2. The 1.5-point difference does not mean the models perform equally on every task.

What does the 82,000-token figure mean?

It is the reported average output-token count per index task for Bolt. The source says that is nearly three times Ultra’s 28,000 tokens per task. It is not a price or a measured completion time.

Is Bolt available through a hosted API, and what does it cost?

The source says Artificial Analysis lists no hosted API provider or price for Bolt. Its weights are available for self-hosting, but actual hardware and serving costs are not given.

Is Bolt the top model in its size class?

No. The source report places Bolt fourth among models under 100 billion total parameters in the cited index comparison, while describing it as first among US models in that class.

Source: ThorstenMeyerAI.com