Nemotron 3.5 Bolt: Ultra-Class Performance In A Smaller Package

🔍 Read the full analysis: Nemotron 3.5 Bolt: Ultra-Class Performance In A Smaller Package on ThorstenMeyerAI.com

Before you orderOffer from Amazon

Get the latest gadgets 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, compared with 22.9 for NVIDIA’s 550-billion-parameter Nemotron 3 Ultra. Bolt has 74 billion total parameters and 9 billion active per token, but it ranks fourth among models under 100 billion parameters and generated substantially more output tokens during evaluation.

The benchmark comparison puts Bolt 1.5 points below Ultra, despite Bolt having about one-seventh as many total parameters and one-sixth as many active parameters. NVIDIA’s earlier Nemotron 3 Super, with 120 billion total parameters and 12 billion active, scored 12.8 in the same index. Artificial Analysis evaluated Bolt’s reasoning version.

Bolt is a mixture-of-experts model with a 256,000-token context window. It accepts text and images and produces text. Its weights are available for self-hosting, and Artificial Analysis classifies the license as permissive, although the model page does not name the specific license. The site lists no hosted API provider, so it has not published a price, speed result or cost-per-task figure.

Results vary by task. Bolt scored 15.5% on AutomationBench-AA, which tests multi-step work across software tools, against Ultra’s 3.0%. But Ultra performed better on long-context evaluation, 79% to Bolt’s 70%, as well as several knowledge-oriented measures. Both models scored near zero on Terminal-Bench 4.0, a result that does not support treating Bolt as a capable shell agent.

Where Bolt Fits in Local AI

Bolt’s main significance is the trade-off it presents for organizations seeking a capable model they can run themselves. Its 9 billion active parameters suggest a lower per-token compute requirement than a larger dense model, although Artificial Analysis has not measured Bolt’s speed, so real-world performance on particular hardware remains unknown.

The model’s standout result is on agentic workflow testing, where it substantially outscored Ultra. That may make it worth testing for tool-based tasks, but the benchmark does not establish that it will perform well in every workflow or production setting. Its weak Terminal-Bench result also cautions against extending the AutomationBench finding to general computer or command-line control.

Token use is another practical consideration. Artificial Analysis reports about 82,000 generated output tokens per index task, compared with 28,000 for Ultra. The source describes this as very verbose relative to models of Bolt’s size. Without measured throughput, the time required to generate that output is unknown. For self-hosting, a smaller parameter count does not by itself mean a faster completed task.

A New Model Below 100B

Bolt is the first release in NVIDIA’s new Bolt model family. It arrived alongside Nemotron 3.5 Super VL, which received more attention at launch. Artificial Analysis places Bolt fourth among models below 100 billion total parameters and first among US models in that group, according to the comparison supplied in the source.

The top of that class is not close on the reported index scores: Qwen3.8 27B scored 33.7, while K2 Horizon MoVA 36B-A4B scored 25.3 and G9v3-39A5B scored 21.8. Bolt’s 21.4 puts it ahead of several other listed models, but below those three. The source notes that Bolt accepts images, unlike K2 Horizon MoVA and G9v3, making the class comparison about more than one score.

The aggregate index also conceals differences among individual tasks. Bolt’s strongest comparison with Ultra is on AutomationBench-AA; Ultra leads on long-context reasoning and some knowledge tests. The reported 256,000-token context window describes how much input the model can accept, not how reliably it reasons across all of that material.

““For batch jobs that run overnight, that is fine. For anything a person waits for, it is not.””

— Thorsten Meyer, ThorstenMeyerAI.com

Speed, Cost and Reliability

Artificial Analysis has not listed a hosted provider for Bolt or published speed, price or cost-per-task measurements. The source’s estimates that 82,000 tokens could take about 14 minutes at 100 tokens per second or 27 minutes at 50 tokens per second are arithmetic examples, not observed Bolt performance. Actual generation time depends on hardware, software and configuration.

The supplied source text ends while discussing Bolt’s negative score on AA-Omniscience, a benchmark that rewards correct answers and penalizes confident errors. It gives a score of −19 but does not provide the rest of the analysis or enough detail to interpret the result fully. The available material also does not establish how Bolt performs across independent deployments, or identify the exact permissive license named by NVIDIA.

As a result, the index score alone cannot settle whether Bolt is preferable to Ultra for a specific use. The benchmark differences, output volume and lack of measured throughput leave real-world latency and reliability open questions.

Awaiting Real-World Measurements

The next useful evidence will be independent speed and throughput tests on the hardware intended for self-hosting, along with evaluations of complete task time rather than scores alone. A hosted API listing, if one appears, could also provide pricing and service-level performance data that are not currently available in the cited model page.

Users assessing Bolt can compare its results on their own workloads, particularly multi-step tool use, long documents and knowledge questions. Those tests should track both answer quality and the number of generated tokens. For now, the confirmed picture is a model close to Ultra on one aggregate index, with a strong result on one agentic benchmark and notable weaknesses and efficiency questions elsewhere.

Key Questions

When did NVIDIA release Nemotron 3.5 Bolt?

NVIDIA released Nemotron 3.5 Bolt 74B-A9B on October 15, 2026, according to the source report published the following day.

How does Bolt compare with Nemotron 3 Ultra?

Bolt scored 21.4 on Artificial Analysis Intelligence Index v4.3.2, versus 22.9 for Nemotron 3 Ultra. Bolt has 74 billion total parameters and 9 billion active per token; Ultra has 550 billion total and 55 billion active.

Is Bolt the top model in its size class?

No. Artificial Analysis ranks it fourth among models under 100 billion parameters in the supplied comparison, though it is listed as the highest-scoring US model in that class.

What tasks does Bolt appear strongest at?

Its reported result is strongest on AutomationBench-AA, a test of multi-step workflows across software tools: 15.5%, compared with 3.0% for Ultra. That result does not mean it performs well on every kind of agent task; both models scored near zero on Terminal-Bench 4.0.

Is Bolt faster or cheaper to run than Ultra?

That has not been established by the cited evaluation. Artificial Analysis lists no hosted API price or speed measurement for Bolt. It also reports higher output-token use per index task than Ultra, which may affect completion time on self-hosted hardware.

Source: ThorstenMeyerAI.com