How Nemotron 3.5 Bolt Trades Model Size For Token Use

🔍 Read the full analysis: How Nemotron 3.5 Bolt Trades Model Size For Token Use on ThorstenMeyerAI.com

Buying for a business?Offer from Amazon

Get business pricing on tech for your team

• 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, an open-weight model that scored 21.4 on the Artificial Analysis Intelligence Index v4.3.2, close to the 22.9 score of NVIDIA’s much larger Nemotron 3 Ultra 550B-A55B. Bolt has 74 billion total parameters, with 9 billion active per token, but its evaluation required about 82,000 output tokens per task—nearly three times Ultra’s reported usage—making speed and practical operating cost central unanswered questions.

The release came on October 15, 2026, alongside Nemotron 3.5 Super VL. Bolt is the first model in NVIDIA’s new Bolt family. Artificial Analysis tested its reasoning version and lists a 256,000-token context window, text-and-image input, text output, and open weights for self-hosting. The source says the model page does not name the license, although Artificial Analysis classifies it as permissive.

Bolt’s index score is 1.5 points below Ultra’s, despite having roughly one-seventh as many total parameters and one-sixth as many active parameters. The result does not make Bolt a leader across all models in its size range: Artificial Analysis ranks it fourth among models under 100 billion parameters, behind Qwen3.8 27B at 33.7, K2 Horizon MoVA 36B-A4B at 25.3, and G9v3-39A5B at 21.8. It is, however, the highest-scoring US model in that group, according to the source.

The benchmark breakdown shows a specific strength in agentic workflows. Bolt scored 15.5% on AutomationBench-AA, which tests multi-step tasks across software services, compared with 3.0% for Nemotron 3 Ultra. It also scored higher on AA-Briefcase and CritPt in the supplied comparison. Ultra did better on several knowledge and document-oriented tests, including AA-LCR, a long-context measure, where it scored 79% to Bolt’s 70%. Both models scored zero on Terminal-Bench 4.0 in the comparison.

A Smaller Model With a Longer Run

Bolt’s result points to a trade-off that matters to teams choosing models for their own infrastructure: parameter count is not the same as task time. A mixture-of-experts design activates 9 billion parameters per token, which could reduce compute needed for each generated token compared with a model that activates far more parameters. But Artificial Analysis has not published a speed measurement for Bolt, and the model generated far more tokens during its index evaluation. The source’s compute-efficiency argument is therefore an architectural expectation, not a measured throughput result.

That token use may affect both cost and responsiveness. For hosted services, additional output can increase charges if billing is token-based; for self-hosting, it can extend the time a GPU is occupied. The source estimates that 82,000 tokens would take about 14 minutes at 100 tokens per second or 27 minutes at 50 tokens per second. Those figures are arithmetic illustrations, not measured Bolt speeds. They show why an index score alone cannot tell users how quickly a model will finish their own tasks.

The workflow results may still interest developers building agents, especially when tasks run unattended or in batches. By contrast, the reported long-document score and lack of a Terminal-Bench result argue against treating a large context window as proof of strong document reasoning or shell-agent ability. Use case and measured latency will matter as much as the headline score.

How Bolt Compares With NVIDIA Models

Bolt follows Nemotron 3 Super, a 120B-total, 12B-active model released in March, which scored 12.8 on the same index, according to the source. Bolt’s 21.4 is an 8.6-point increase over that result, while using fewer total parameters. The comparison signals a sharp improvement within NVIDIA’s own open-weight lineup, but scores from one index do not establish that Bolt is better for every task.

The comparison with Ultra is also a comparison of different profiles, not simply a smaller version of the same system. In the supplied results, Bolt leads on AutomationBench-AA, while Ultra scores higher on long-context reasoning and several knowledge-oriented evaluations. On AA-Omniscience, which rewards correct answers and penalizes confident wrong ones, the source reports Bolt at minus 19 and Ultra at zero. It also notes that neither model performed usefully on Terminal-Bench 4.0.

Within the under-100B group, Bolt’s score trails Qwen3.8 27B by 12.3 points and K2 Horizon MoVA by 3.9. The source says Bolt accepts images, unlike K2 Horizon MoVA and G9v3, and may interest organizations that need to document US model provenance. Those are relevant selection factors, but they do not change its fourth-place ranking or establish comparative deployment speed.

“Bolt’s ‘Ultra-class score’ is an Ultra-class score per task, not per minute.”

— Thorsten Meyer, ThorstenMeyerAI.com

Speed, Pricing and License Details

No speed or cost-per-task figure is available in the source because Artificial Analysis lists no hosted API provider for Bolt and has not measured its speed. The token-count estimates cannot substitute for real-world throughput: hardware, serving setup, prompt length, and workload can affect completion time. The source also does not establish what Bolt would cost through a hosted service.

Several practical points remain unsettled: the model page reportedly does not identify the exact permissive license; benchmark results do not show how the model will perform across users’ specific agent workflows; and its image-input capability is listed without details here about image-task evaluation. The source’s material ends during its discussion of Bolt’s knowledge and hallucination results, so no further evidence or conclusions on those points are available in the supplied account.

Independent Tests Will Show Deployment Trade-Offs

The next useful evidence will be measured inference speed and real deployment costs on specified hardware, alongside details of any hosted API and its pricing. Those measurements would test whether Bolt’s low active-parameter count offsets its high token use in practice. Users evaluating it for agents will also need task-specific comparisons, rather than relying on one benchmark score.

Teams considering self-hosting can review the released weights and verify the license terms before deployment. Until providers publish performance data and independent users report results on their own workloads, Bolt’s strongest confirmed case is its index score and AutomationBench-AA result—not a demonstrated advantage in response time or cost.

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 and activates 9 billion per token, according to the supplied report.

How close is Bolt to Nemotron 3 Ultra?

Bolt scored 21.4 on Artificial Analysis Intelligence Index v4.3.2, versus 22.9 for Nemotron 3 Ultra 550B-A55B. That is a 1.5-point gap on that index, not proof of equal performance across all tasks.

Why is Bolt’s token use a concern?

The source reports about 82,000 output tokens per index task, nearly three times Ultra’s 28,000. More generated tokens can mean longer waits on local hardware or higher charges where hosted services bill for output; Bolt’s actual speed and hosted price are not reported.

Is Bolt the top model under 100 billion parameters?

No. Artificial Analysis ranks it fourth in the under-100B group in the supplied account. It is reported as the highest-scoring US model in that class, while Qwen3.8 27B, K2 Horizon MoVA 36B-A4B, and G9v3-39A5B score higher overall.

What should users wait to learn?

Users still need published speed tests, deployment costs, and exact license details. Independent evaluations on specific workloads would also help establish whether Bolt’s agent-task results translate into useful performance outside the index.

Source: ThorstenMeyerAI.com