Nemotron 3.5 Bolt: Ultra-Class Scores At One-Seventh The Size, With A Token Cost

🔍 Read the full analysis: Nemotron 3.5 Bolt: Ultra-Class Scores At One-Seventh The Size, With A Token Cost 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 Artificial Analysis’s Intelligence Index v4.3.2, close to the 22.9 score of the company’s 550B-parameter Nemotron 3 Ultra. The result comes with a trade-off: Artificial Analysis’s evaluation recorded about 82,000 output tokens per index task, and Bolt ranks fourth rather than first among models under 100B parameters.

Bolt is a mixture-of-experts model with 74 billion total parameters and 9 billion active for each token, according to the source report. It has a 256,000-token context window, accepts text and images, and generates text. NVIDIA released it with open weights for self-hosting. Artificial Analysis evaluated the reasoning version.

The comparison with Ultra is close on the headline index score: Bolt is 1.5 points behind, despite having about one-seventh as many total parameters and one-sixth as many active parameters. The report says Bolt scored 8.6 points above NVIDIA’s Nemotron 3 Super, a 120B-total, 12B-active model released in March. These figures describe benchmark results, not speed or cost in everyday use.

Artificial Analysis lists Bolt as fourth among models under 100B, and first among US models in that size class. Its score of 21.4 trails Qwen3.8 27B at 33.7, K2 Horizon MoVA 36B-A4B at 25.3 and G9v3-39A5B at 21.8. The ranking and scores are reported by the source; they do not establish that Bolt is best for every task.

Where Bolt Could Fit

Bolt’s most striking result in the supplied comparisons is on AutomationBench-AA, which tests multi-step workflows across software tools. It scored 15.5%, compared with 3.0% for Nemotron 3 Ultra and 14.8% for Nemotron 3.5 Super VL, according to the report. That suggests Bolt may be worth testing for agentic workflows, though the score alone does not establish dependable performance in production.

The model’s image input and open weights may also matter to organizations that need to run models on their own infrastructure or document model provenance. The report says Bolt is the highest-scoring US model in its under-100B category. For self-hosting, its 9B active parameters may reduce computation per generated token compared with a larger dense model, but Artificial Analysis has not published a speed measurement for Bolt. That potential hardware advantage remains unverified.

Token volume is a separate practical concern. Artificial Analysis recorded 82,000 output tokens per index task, nearly three times the 28,000 reported for Ultra. At a hypothetical generation rate of 100 tokens per second, 82,000 tokens would take about 14 minutes to produce; at 50 tokens per second, about 27 minutes. These are calculations, not measured Bolt runtimes, and actual throughput depends on hardware and deployment.

Benchmark Strengths and Gaps

The Intelligence Index score does not capture all of Bolt’s strengths and weaknesses. In the supplied comparison, Bolt scored 15.5% on AutomationBench-AA, but it lagged Ultra on several other measures: 70% versus 79% on AA-LCR, a long-context reasoning test, and 25% versus 28% on Humanity’s Last Exam. Ultra also scored higher on GDPval-AA and GDP.pdf. On SciCode, the two were close, at 39.7% for Bolt and 40.3% for Ultra.

The report says neither model performed usefully on Terminal-Bench 4.0, with Bolt at 0% and Ultra at 0.5%. That limits the case for treating Bolt as a shell-operating agent. Its large context window is a specification, not proof that it can reliably reason across every document placed within that limit.

Artificial Analysis also gave Bolt a score of −19 on AA-Omniscience, a test that rewards correct answers and penalizes confident wrong ones. The source excerpt ends while discussing this result, so it does not provide further explanation or a full interpretation. Readers should not infer a specific error rate from that score alone.

“Bolt is not a shrunken Ultra. It is better at agentic work and weaker at knowledge and long documents.”

— Thorsten Meyer, ThorstenMeyerAI.com

Pricing, Speed and License Details

Artificial Analysis lists no hosted API provider for Bolt, so the source report gives no API price, measured speed or cost per task. That leaves the practical trade-off between quality, latency and expense unresolved. The 14- and 27-minute examples are theoretical estimates based on assumed token-generation rates, not observed performance.

The report describes Bolt’s license as permissive, but says the model page does not name the license. It also does not provide independent testing beyond the cited evaluations, or enough information to determine how performance will vary across hardware and real-world workloads. Those points should be checked against NVIDIA’s model materials before deployment.

What Deployers Should Verify

The next useful evidence will be measured inference speed, hosted-provider availability and pricing, plus the exact license terms. Until those details are available, teams considering self-hosting can test Bolt on their own hardware and tasks, recording output length, latency and quality rather than relying on the Intelligence Index score alone.

Further independent evaluations could clarify whether Bolt’s AutomationBench result carries over to real software workflows and how its weaker long-context and knowledge scores affect other uses. For now, the confirmed picture is a compact open-weight model with a near-Ultra index score, a strong result on one agent benchmark, and unusually high token use in the cited evaluation.

Key Questions

What is Nemotron 3.5 Bolt?

Nemotron 3.5 Bolt 74B-A9B is NVIDIA’s open-weight mixture-of-experts model, released October 15, 2026. It has 74 billion total parameters, 9 billion active per token, accepts text and images, and outputs text.

How close did Bolt score to Nemotron 3 Ultra?

Artificial Analysis’s Intelligence Index v4.3.2 gave Bolt 21.4 and Nemotron 3 Ultra 22.9, a 1.5-point difference. This is a benchmark comparison, not a direct measure of speed or cost.

Is Bolt the top model under 100 billion parameters?

No. The supplied report places Bolt fourth in the under-100B category and first among US models in that class. Three listed models scored higher on the index.

Why does Bolt’s token use matter?

Artificial Analysis recorded about 82,000 output tokens per index task. More generated tokens can mean longer waits on self-hosted systems and potentially higher charges from a paid API, but Bolt’s actual speed and hosted cost are not yet listed in the supplied material.

Source: ThorstenMeyerAI.com