Nemotron 3.5 Bolt: How Its Size And Token Trade-Off Compare

🔍 Read the full analysis: Nemotron 3.5 Bolt: How Its Size And Token Trade-Off Compare 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 Intelligence Index v4.3.2, 1.5 points below the company’s 550B-parameter Nemotron 3 Ultra. Bolt has 74 billion total parameters and activates 9 billion per token, but it ranks fourth among models under 100B parameters and generated about 82,000 output tokens per index task—a potential time cost for people running it on their own hardware.

The model uses a mixture-of-experts architecture, supports text and image input, produces text, and has a 256,000-token context window. NVIDIA released it with open weights for self-hosting. Artificial Analysis evaluated its reasoning version. The evaluator’s model page lists no hosted API provider, price, speed measurement, or cost per task; it also classifies the license as permissive without naming the license on the page.

Against Nemotron 3 Ultra, Bolt’s index score is close despite a much smaller model: 74B total parameters versus 550B, and 9B active parameters versus 55B. The result is not a uniform improvement. In Artificial Analysis’s reported comparisons, Bolt scored higher on AutomationBench-AA (15.5% versus 3.0%) and AA-Briefcase (Elo 912 versus 876). Ultra scored higher on long-context AA-LCR (79% versus 70%), Humanity’s Last Exam (28% versus 25%), and the GDP.pdf test (5.0% versus 2.8%).

The evaluator places Bolt fourth among models under 100B total parameters and first among US models in that group. The listed class leader, Qwen3.8 27B, scored 33.7; K2 Horizon MoVA 36B-A4B scored 25.3, and G9v3-39A5B scored 21.8. NVIDIA’s earlier Nemotron 3 Super, with 120B total and 12B active parameters, scored 12.8 on the index. Artificial Analysis’s comparisons show Bolt’s strength on a particular agent-workflow test, but do not establish that it is the strongest model overall for agent use.

Smaller Model, Longer Inference

Bolt’s score puts a near-flagship result within a model with far fewer total and active parameters. For organizations that self-host models, that architecture may be relevant to hardware planning. However, fewer active parameters do not by themselves establish faster inference or lower end-to-end cost: Artificial Analysis has not published a speed measurement for Bolt, and its token output is unusually high.

Artificial Analysis reports that Bolt generated about 230 million output tokens across its Intelligence Index evaluation, describing it as “very verbose” compared with a 100 million-token median for open-weight models of its size. Its per-task output was about 82,000 tokens, compared with 28,000 for Nemotron 3 Ultra. Those figures suggest that users may wait longer for results when running Bolt locally, but actual runtime depends on hardware, inference software, settings, and workload.

The difference matters most when people interact with the model in real time. For overnight or batch processing, additional generated tokens may be acceptable; for a person waiting on an answer, tokens per task can matter as much as a benchmark score. At hypothetical generation rates of 100 and 50 tokens per second, 82,000 tokens would take roughly 14 and 27 minutes, respectively. These are arithmetic illustrations, not measured Bolt runtimes.

Where Bolt Fits in NVIDIA’s Lineup

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 report says Super VL drew more attention, while Bolt stood out for its benchmark score relative to its size. The comparison with Nemotron 3 Ultra is against NVIDIA’s previous flagship, which scored 22.9 on the same cited index version.

Bolt’s results vary by task. Artificial Analysis’s AutomationBench-AA measures multi-step workflows across software tools; Bolt scored 15.5%, compared with Ultra’s 3.0%. That is a notable result for agent-style workflows, but it should not be read as general competence across all tools or tasks. On Terminal-Bench 4.0, the supplied figures are 0% for Bolt and 0.5% for Ultra, while the long-document AA-LCR score favors Ultra. The benchmarks therefore describe different strengths and limitations rather than a single ranking.

Within the under-100B group, Bolt’s 21.4 is well below Qwen3.8 27B’s 33.7 and K2 Horizon MoVA’s 25.3. The source report also notes that Bolt accepts images, while those two comparison models do not, and that its 9B active parameters are an architectural distinction. These features may matter to buyers with particular deployment or input needs, but Artificial Analysis has not measured Bolt’s speed to verify a practical performance advantage.

““For a 9B-active model, that is a real result.””

— Thorsten Meyer, ThorstenMeyerAI.com

Speed, Price and License Details

No measured inference speed or hosted price is listed in the supplied Artificial Analysis model-page details. The token count indicates how much Bolt generated in that evaluation, but it does not determine how long a given user’s setup will take. Hardware, software, decoding settings, and task requirements can all affect runtime.

The model page classifies the weights’ license as permissive, but the source report says it does not identify the license by name. The practical terms for a particular deployment should not be inferred from the category alone. It is also unclear from the supplied material how the benchmark results change under different prompting or inference settings, or whether later model versions will reduce token use.

The index score is one benchmark result, not evidence that Bolt matches Ultra across tasks. Bolt’s lower AA-LCR and knowledge-related results, as reported in the source, point to meaningful differences. Artificial Analysis also assigns Bolt a score of −19 on AA-Omniscience, a benchmark that rewards correct answers and penalizes confident errors; the supplied source text cuts off before giving further interpretation, so the result should not be expanded into a broader claim about reliability.

Awaiting Runtime and License Data

The clearest next comparison will be measured inference speed and cost on specified hardware, alongside token use on comparable tasks. Those results would show whether Bolt’s smaller active parameter count translates into faster or more economical self-hosting despite its longer outputs. A hosted API listing, if one becomes available, could also add published pricing and speed data.

Users considering deployment can already compare the reported benchmark profile with their own requirements: image input and software-workflow performance may be relevant, while long-document reasoning, knowledge tests, and terminal use require closer scrutiny. The named license and its terms also remain a practical point to verify. Until further measurements or documentation are published, Bolt’s near-Ultra index score is confirmed in the cited evaluation, but its real-world speed and cost advantage remain unestablished.

Key Questions

What is Nemotron 3.5 Bolt?

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

How does 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 count, but the models have different benchmark strengths: Bolt scored higher on AutomationBench-AA, while Ultra scored higher on long-context AA-LCR and several other listed tests.

Is Bolt the top model in its size class?

No. Artificial Analysis ranks Bolt fourth among models under 100B total parameters in the supplied comparison. Qwen3.8 27B leads that list at 33.7, followed by K2 Horizon MoVA 36B-A4B at 25.3 and G9v3-39A5B at 21.8.

Why could Bolt take a long time to run?

Artificial Analysis reports that Bolt generated about 82,000 output tokens per index task, nearly three times Nemotron 3 Ultra’s 28,000. That may mean longer waits on some systems, but the source provides no measured speed; actual runtime depends on hardware and software.

Can users access Bolt through a hosted API?

The supplied Artificial Analysis model-page information lists no hosted API provider, price, or speed measurement. It describes the weights as open for self-hosting and classifies the license as permissive, without naming the license.

Source: ThorstenMeyerAI.com