Nemotron 3.5 Bolt: An Ultra-Class Score At A Seventh Of The Size. You Pay For It In Tokens.

🔍 Read the full analysis: Nemotron 3.5 Bolt: An Ultra-Class Score At A Seventh Of The Size. You Pay For It In Tokens. on ThorstenMeyerAI.com

Before you orderOffer from Amazon

Get movie nights 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 October 15, an open-weight model that scored 21.4 on Artificial Analysis Intelligence Index v4.3.2, close to the company’s 550B-parameter Nemotron 3 Ultra, which scored 22.9. Bolt has about one-seventh of Ultra’s total parameters, but the comparison comes with qualifications: it ranks fourth among models below 100B parameters and generated far more output tokens during evaluation.

Bolt uses a Mixture of Experts architecture with 74 billion total parameters and 9 billion active per token. NVIDIA lists a 256,000-token context window, text and image input, text output, and reasoning capability. The model’s weights are available for self-hosting. Artificial Analysis classifies its license as permissive, although the model page does not name the license.

Artificial Analysis measured Bolt at 21.4 points, against 22.9 for Nemotron 3 Ultra 550B-A55B, which has 550 billion total and 55 billion active parameters. Bolt’s score is 1.5 points lower, with roughly one-sixth as many active parameters. NVIDIA’s earlier Nemotron 3 Super, a 120B-total, 12B-active model released in March, scored 12.8 on the same index, according to the source report.

The benchmark profile is mixed. Artificial Analysis results cited by ThorstenMeyerAI.com show Bolt scoring 15.5% on AutomationBench-AA, against 3.0% for Ultra. The test covers multi-step workflows across software tools. Bolt also led Ultra on AA-Briefcase and CritPt in the reported comparison. Ultra scored higher on long-context reasoning, knowledge-oriented tests and several other measures, including AA-LCR, where it earned 79% to Bolt’s 70%. Both models scored near zero on Terminal-Bench 4.0.

A Smaller Model With a Long Wait

Bolt’s result matters to developers deciding whether a smaller self-hosted model can handle demanding work. Its index score comes close to NVIDIA’s much larger Ultra, and its reported AutomationBench advantage suggests potential for some multi-step software workflows. The model also accepts images, a feature the source says is absent from two higher-scoring sub-100B competitors, K2 Horizon MoVA and G9v3.

Parameter counts alone do not establish real-world speed or cost. Bolt generated about 82,000 output tokens per index task, compared with about 28,000 for Ultra, according to the cited evaluation. The article’s estimates of roughly 14 minutes at 100 tokens per second and 27 minutes at 50 tokens per second are arithmetic illustrations, not measured Bolt throughput. They show why a strong score per task may not translate into a fast response for someone waiting on an answer.

For teams running batch jobs on their own hardware, longer generation may be acceptable if the model’s capability and operating requirements fit their workload. For interactive use, the absence of published speed measurements makes the practical trade-off harder to judge. Organizations that track model origin may also value that Artificial Analysis places Bolt first among US models in its under-100B category, though not first overall.

Bolt Among Models Under 100B

Artificial Analysis ranks Bolt fourth among models with fewer than 100 billion total parameters in the comparison reported by ThorstenMeyerAI.com. Qwen3.8 27B leads the listed group at 33.7, followed by K2 Horizon MoVA 36B-A4B at 25.3 and G9v3-39A5B at 21.8. Bolt scored 21.4, ahead of K2 Horizon 7B at 20.6 and the other models listed in the source.

That ranking puts the near-match with Ultra in perspective: Bolt is not the top-scoring model in its size group. The report says it is the highest-scoring US model below 100B, nearly four points ahead of Meta’s Muse Glimmer 30B. The source also notes that K2 Horizon MoVA scores higher while using fewer total and active parameters; its reported index score is 25.3.

Artificial Analysis has not listed a hosted API provider for Bolt, according to the source. As a result, there is no published hosted price, speed measurement or cost-per-task figure in the cited model listing. The comparison is based on benchmark results and reported token output, not a complete deployment-cost test.

“The headline is a gap of 1.5 points, from a model with about a seventh of the total parameters.”

— Thorsten Meyer, ThorstenMeyerAI.com

Speed, Cost and License Details

Several practical details remain unavailable. Artificial Analysis has no hosted API provider listed for Bolt, so its hosted price and measured inference speed are unknown. The cited 14- and 27-minute examples assume fixed token-generation rates and should not be read as test results. Actual performance will depend on hardware, software and deployment configuration.

The source identifies the license as permissive but says the model page does not specify its name. It also reports a negative score of −19 on AA-Omniscience, a test that rewards correct answers and penalizes confident errors, but the supplied material ends before explaining the result further. The reasons for Bolt’s large output volume and how that affects different prompts are not fully established by the cited information.

Deployment Tests Will Settle Trade-Offs

The next useful evidence will be independent throughput and cost testing across common hardware, along with details of any hosted API options and the precise model license. Those measurements would show whether Bolt’s lower active-parameter count translates into practical efficiency despite its high token use.

Users evaluating the model can compare it against their own tasks rather than relying on the index score alone. In particular, the reported strength on tool-based automation should be weighed against weaker long-context results and the lack of useful performance on Terminal-Bench 4.0 in the cited comparison. No additional NVIDIA release or evaluation timetable was provided in the source material.

Key Questions

What is Nemotron 3.5 Bolt?

It is an open-weight NVIDIA model released October 15, 2026. It has 74 billion total parameters, 9 billion active per token, supports text and image input, and produces text.

How close is Bolt to Nemotron 3 Ultra?

Artificial Analysis scored Bolt 21.4 and Nemotron 3 Ultra 550B-A55B 22.9 on Intelligence Index v4.3.2. Bolt has about one-seventh of Ultra’s total parameters, but the score does not establish equal performance on every task.

Is Bolt the top model under 100B parameters?

No. The source report says Artificial Analysis ranks it fourth in the under-100B group and first among US models in that category.

How fast or expensive is Bolt to run?

No hosted price or measured speed is listed in the cited material. Bolt generated about 82,000 output tokens per index task; estimates of how long that takes at assumed token rates are calculations, not benchmark measurements.

What kinds of tasks appear to suit Bolt?

The reported results point to a relative strength in multi-step tool workflows: Bolt scored 15.5% on AutomationBench-AA, compared with 3.0% for Ultra. It scored lower than Ultra on long-context reasoning and several knowledge-oriented measures, so performance depends on the task.

Source: ThorstenMeyerAI.com