Nemotron 3.5 Bolt: Strong Scores From A Smaller Model, At A Token Cost

🔍 Read the full analysis: Nemotron 3.5 Bolt: Strong Scores From A Smaller Model, At A Token Cost on ThorstenMeyerAI.com

Buying for a business?Offer from Amazon

Get business pricing on office and shipping supplies

• Business-only prices and quantity discounts
• Tax-exempt purchasing
• Multiple users, one account, clear invoices

Create a free Business account →Free to register

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 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 the benchmark report also found Bolt used about 82,000 output tokens per task—nearly three times Ultra’s count.

The index score puts Bolt close to NVIDIA’s previous flagship on that benchmark, despite its substantially smaller parameter count. Against NVIDIA’s Nemotron 3 Super, released in March with 120 billion total parameters and 12 billion active, Bolt scored 8.6 points higher while using 39% fewer total parameters, according to the figures reported by ThorstenMeyerAI.com. Those comparisons describe benchmark results; they do not establish that Bolt is faster or cheaper to run.

Bolt ranks fourth among models under 100 billion parameters in Artificial Analysis’s comparison, and is the highest-ranked US model in that size group. The reported 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. The score places Bolt ahead of several other listed models, but well behind Qwen3.8.

The evaluation describes a mixed capability profile rather than a smaller replica of Ultra. Bolt scored 15.5% on AutomationBench-AA, compared with Ultra’s 3.0%, a benchmark of multi-step workflows across software tools. But Bolt scored 70% on AA-LCR, a long-context reasoning test, versus Ultra’s 79%. The source also reports that neither model scored usefully on Terminal-Bench 4.0: Bolt recorded 0% and Ultra 0.5%.

A Smaller Model With a Longer Run

Bolt’s results matter to teams weighing open-weight models for self-hosting because the model combines a strong score on a prominent index with a relatively small active parameter count. Its 9 billion active parameters per token could make it less demanding to run than larger models in some setups. That is an architectural possibility, not a measured performance result: Artificial Analysis has not published a speed measurement for Bolt, and the report lists no hosted API provider or price.

The token count complicates any claim of efficiency. Artificial Analysis reported about 230 million output tokens across the index evaluation, describing Bolt as very verbose relative to its size group. Its reported average of roughly 82,000 output tokens per task is about 2.9 times Ultra’s 28,000. More generated tokens can mean longer waits on local hardware and, if offered through a paid API, potentially higher usage charges. No API pricing is available to quantify that cost.

For prospective users, the benchmark differences may matter more than the headline index gap. The AutomationBench result suggests Bolt may merit testing for agentic workflows, while its lower long-document score and reported weakness on an omniscience measure caution against assuming broad superiority. The results are benchmark-specific; they do not establish how Bolt will perform on a particular organization’s data or tasks.

Bolt in NVIDIA’s Model Line

NVIDIA announced two open-weight models on October 15: Nemotron 3.5 Super VL and Bolt, the first model in a new Bolt family. The source says Super VL drew more attention, while presenting Bolt as the notable engineering result because of its score relative to its parameter count. It describes Super VL as the highest-scoring permissively licensed open-weight model from the United States, a claim specific to the source’s comparison and timing.

Bolt accepts text and images and produces text. It supports a 256,000-token context window and has a reasoning version, which Artificial Analysis evaluated. The model is available with open weights for self-hosting. Artificial Analysis classifies its license as permissive, though the license name is not listed on the model page cited in the report.

The size-class comparison highlights that Bolt’s significance is not an overall lead. Qwen3.8 27B is reported at 33.7, 12.3 points above Bolt, while K2 Horizon MoVA scores 25.3 with 36 billion total parameters and four billion active. The source points to Bolt’s US provenance and image input as distinctions, alongside its architecture. Its reported ranking and the relevance of provenance depend on the comparison set and the requirements of the organization choosing a model.

“Bolt is fourth in its own size class, not first.”

— Thorsten Meyer, ThorstenMeyerAI.com

Speed, Cost and Reliability Gaps

There is no published speed measurement or hosted API price in the cited model listing, so the practical latency and cost of Bolt remain unknown. The report gives illustrative arithmetic: at 100 tokens per second, 82,000 tokens would take about 14 minutes to generate; at 50 tokens per second, about 27 minutes. Those are calculations based on assumed speeds, not measurements of Bolt on specific hardware. Real results would depend on the system, configuration and workload.

The source material ends during its discussion of Bolt’s knowledge and hallucination performance. It states that Bolt scored −19 on Artificial Analysis’s AA-Omniscience measure, which rewards correct answers and penalizes confident wrong answers, but does not provide the rest of the interpretation or enough detail to describe the result further. The full implications of that score are therefore unclear from the material provided. Benchmark scores also do not establish performance across all real-world uses.

The license is another detail needing verification by users: the report says Artificial Analysis categorizes it as permissive but does not name the license on the model page. Teams considering deployment should check the model’s published license and release documentation rather than infer its terms from the classification alone.

Independent Tests Will Settle Practicality

The next useful evidence will be measured inference speed and task-level results on the hardware and workloads users actually plan to run. A hosted provider listing, if one appears, could supply API pricing and response-time comparisons; none is listed in the source report. Independent tests could also show whether Bolt’s token-heavy reasoning is manageable for batch jobs and whether it creates delays in interactive use.

Users evaluating Bolt should compare it with alternatives on the tasks they need, including long-document reasoning and tool use, rather than relying on the aggregate index alone. Further reporting may also clarify the AA-Omniscience result and the model’s license terms. Until then, the confirmed picture is a close-to-Ultra index score from a smaller model, paired with a high output-token count and no published speed or price measurement.

Key Questions

What is Nemotron 3.5 Bolt?

Nemotron 3.5 Bolt 74B-A9B is an open-weight NVIDIA model released October 15, 2026. It has 74 billion total parameters, 9 billion active per token, accepts text and images, and supports a 256,000-token context window.

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 is much smaller by total and active parameter counts, but the source reports that it generated roughly 82,000 output tokens per index task versus Ultra’s 28,000.

Is Bolt the top model in its size class?

No. The report places Bolt fourth among models under 100 billion parameters, behind Qwen3.8 27B, K2 Horizon MoVA 36B-A4B and G9v3-39A5B. It is reported as the highest-ranked US model in that class.

Is Bolt fast or inexpensive to run?

That has not been established by the cited information. Artificial Analysis has no listed speed measurement or hosted API provider for Bolt, and no API price is given. The reported token count may mean longer local generation times, but actual latency depends on hardware and configuration.

What tasks appear to suit Bolt, based on these benchmarks?

Bolt scored higher than Nemotron 3 Ultra on AutomationBench-AA, but lower on the long-context AA-LCR test. These results suggest a mixed profile, not a general recommendation; users should test the model on their own workflows and verify its limitations.

Source: ThorstenMeyerAI.com