A Smaller AI Model With An Ultra-Class Score: The Nemotron 3.5 Bolt Trade-Off

🔍 Read the full analysis: A Smaller AI Model With An Ultra-Class Score: The Nemotron 3.5 Bolt Trade-Off 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 on October 15, 2026, an open-weight model that scored 21.4 on Artificial Analysis’s Intelligence Index v4.3.2, close to the 22.9 recorded by NVIDIA’s 550B-parameter Nemotron 3 Ultra. Bolt has 74 billion total parameters and 9 billion active per token, but it ranked fourth among models under 100 billion parameters and generated about 82,000 output tokens per index task, a trade-off that could mean longer waits when running it locally.

Artificial Analysis evaluated Bolt’s reasoning version. NVIDIA describes the model as open-weight and suitable for self-hosting; it accepts text and images, produces text, supports reasoning, and has a 262,144-token context window. The model page lists no hosted API provider, so there is no reported API price, speed measurement, or cost per task. Artificial Analysis classifies its license as permissive, although the model page does not name the license.

The overall score is not a like-for-like win across every capability. Bolt scored 15.5% on AutomationBench-AA, which evaluates multi-step workflows across SaaS tools; Nemotron 3 Ultra scored 3.0%. Bolt also led Ultra on AA-Briefcase and CritPt, while Ultra scored higher on GDPval-AA, long-context reasoning, Humanity’s Last Exam, and GDP.pdf. Both scored close to zero on Terminal-Bench 4.0, so the supplied results do not support treating either model as a capable shell agent.

Artificial Analysis reports that Bolt used about 82,000 output tokens per index task, compared with 28,000 for Ultra. That is a measure of generated tokens, not a direct speed test. At a hypothetical generation rate of 100 tokens per second, 82,000 tokens would take about 14 minutes; at 50 tokens per second, about 27 minutes. Those are arithmetic estimates, not measured Bolt performance.

A Smaller Model, With Longer Runs

Bolt’s benchmark result is relevant to organizations weighing model size, capability, and deployment control. Its score is 1.5 points below Nemotron 3 Ultra despite having roughly one-seventh the total parameters and one-sixth the active parameters. That suggests NVIDIA has produced a much smaller model with a broadly comparable aggregate score, though the index does not mean the models perform equally on every task.

For agent workflows, the AutomationBench-AA result is a particular strength: Bolt’s 15.5% was well above Ultra’s 3.0% in the supplied comparison. But the extra output tokens complicate the practical case. For a self-hosted deployment, longer generations can consume GPU time and make interactive use slower; for a hosted service, output volume may affect billing, though Bolt has no listed provider or price at this point.

Its ranking also puts the headline in perspective. Bolt is fourth in the under-100B group, not the leading model overall in that size class. The source comparison lists Qwen3.8 27B at 33.7 and K2 Horizon MoVA 36B-A4B at 25.3. Bolt’s reported distinction is that it is the highest-scoring US model in the class, a point that may matter to organizations documenting model origin and provenance. Its image input is another difference from some higher-scoring models in the comparison.

NVIDIA’s Smaller-Model Progress

Bolt is the first model in NVIDIA’s new Bolt family, according to the supplied report. It follows Nemotron 3 Super, a 120B-total, 12B-active model released in March that scored 12.8 on the same cited index. The report calculates that Bolt’s score is 8.6 points higher while using 39% fewer total parameters. Those comparisons describe benchmark results, not a guarantee of better results across all uses.

NVIDIA’s previous flagship, Nemotron 3 Ultra 550B-A55B, provides the reference point for the near-match claim. Bolt’s architecture is a mixture of experts, with 74 billion total parameters and 9 billion active for each token. That active-parameter figure may imply lower compute needs than a dense model of similar total size, but the supplied material includes no measured inference speed for Bolt. Any hardware-efficiency conclusion remains unverified until such results are available.

Artificial Analysis’s index combines multiple evaluations, and the listed task results show why a single score needs qualification. Bolt performs much better than Ultra on the cited automation benchmark, while Ultra is ahead on long-context and several knowledge-oriented tasks. Bolt’s 256K context window describes how much input it can accept; it does not by itself show how reliably it can reason across that input.

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

— Thorsten Meyer, ThorstenMeyerAI.com

Speed, Cost, and Knowledge Gaps

Real-world inference speed remains unreported in the supplied material. The token-based timing examples are estimates at assumed generation rates, not benchmark measurements, and do not account for hardware, batching, prompt length, or other system effects. It is not yet clear how quickly Bolt will run on commonly used self-hosting setups.

The source also provides no hosted API availability or price, and the license name is not specified on the model page. The supplied excerpt ends while discussing Bolt’s negative AA-Omniscience score, so it does not provide the full explanation or related figures needed to characterize that result. The available comparisons indicate weaker performance than Ultra on several knowledge and long-document measures, but do not establish why those gaps occur or how they translate to specific deployments.

Finally, benchmark scores are evaluation results, not guarantees of production performance. The results establish that Bolt ranked fourth in the cited under-100B comparison at the time reported; rankings may change as more models or configurations are evaluated.

Testing Bolt in Real Deployments

The next useful evidence will be independent speed tests on specified hardware, alongside measurements of output length, task success, and resource use. Those results would help users judge whether Bolt’s smaller active-parameter count offsets its longer generations in practical workloads.

Organizations considering self-hosting can also check NVIDIA’s model documentation for the exact license terms and deployment requirements, and compare performance on their own tasks rather than relying on the index score alone. If a hosted API becomes available, its pricing and measured latency will add information that is currently missing. Until then, Bolt’s strongest case in the supplied results is agentic workflow performance; its efficiency and interactive responsiveness remain open questions.

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 produces text.

How close is Bolt to Nemotron 3 Ultra?

Bolt scored 21.4 on Artificial Analysis’s Intelligence Index v4.3.2, compared with 22.9 for Nemotron 3 Ultra 550B-A55B. The aggregate scores are close, but the models differ across individual evaluations.

Is Bolt the top model under 100 billion parameters?

No. The supplied Artificial Analysis comparison ranks it fourth in that group. The report identifies it as the highest-scoring US model in the class, while listing higher scores for Qwen3.8 27B, K2 Horizon MoVA 36B-A4B, and G9v3-39A5B.

Why might Bolt take longer to use?

Artificial Analysis reports about 82,000 output tokens per index task, almost three times Nemotron 3 Ultra’s 28,000. That suggests more generated text, but actual completion time depends on the hardware and generation speed; the source gives no measured Bolt speed.

Can Bolt be used as a shell agent?

The supplied results do not support that use. Both Bolt and Nemotron 3 Ultra scored close to zero on Terminal-Bench 4.0. Bolt’s stronger result was on AutomationBench-AA, a separate evaluation of multi-step workflows across SaaS tools.

Source: ThorstenMeyerAI.com