The Upside And Token Cost Of Nemotron 3.5 Bolt’s Smaller Size

🔍 Read the full analysis: The Upside And Token Cost Of Nemotron 3.5 Bolt’s Smaller Size on ThorstenMeyerAI.com

Buying for a business?Offer from Amazon

Get business pricing on monitors, keyboards and dev gear

• 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 Oct. 15, and an Artificial Analysis evaluation scored it 21.4 on the Intelligence Index, within 1.5 points of the company’s 550B-parameter Nemotron 3 Ultra. The smaller open-weight model 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 may affect the time it takes to run locally.

The Artificial Analysis Intelligence Index version cited in the source is v4.3.2. Bolt scored 21.4, compared with 22.9 for Nemotron 3 Ultra 550B-A55B, which has about seven times as many total parameters and six times as many active parameters. NVIDIA’s Nemotron 3 Super, released in March with 120B total and 12B active parameters, scored 12.8 on the same index, according to the source. These scores indicate performance on that evaluation; they do not establish that the models are interchangeable across tasks.

Bolt’s results vary by task. Artificial Analysis recorded a 15.5% score on AutomationBench-AA, compared with 3.0% for Ultra, and Bolt also scored higher on AA-Briefcase. Ultra performed better on long-context reasoning, with 79% on AA-LCR versus Bolt’s 70%, and on several knowledge-related or document tasks listed in the comparison. Both models scored near zero on Terminal-Bench 4.0, so the source cautions against treating Bolt as a capable shell agent.

The model accepts text and images, produces text, supports reasoning and has a listed 256,000-token context window. Its weights are available for self-hosting. Artificial Analysis had no hosted API provider listed in the cited material, leaving no published provider price, speed measurement or cost-per-task figure there. The source describes the license classification as permissive but says the model page did not identify the license by name.

Token Use Changes Bolt’s Local Trade-Off

Bolt’s smaller active parameter count may make it attractive for organizations seeking an open-weight model that can run on their own infrastructure. It also scored strongly on the cited agentic-work benchmark: 15.5% on AutomationBench-AA, which tests multistep work across software tools. That result is relevant to teams exploring workflow agents, but it should not be read as proof of broad reliability; the benchmark score is low in absolute terms and does not tell users how the model will perform on their specific systems.

The evaluation’s token count complicates the efficiency case. Artificial Analysis reports roughly 82,000 output tokens per index task, versus 28,000 for Ultra. The source’s arithmetic illustrates possible generation time: at a sustained 100 tokens per second, 82,000 tokens would take about 14 minutes; at 50 tokens per second, about 27 minutes. Those are estimates, not measured Bolt run times, and actual throughput depends on hardware and configuration. Still, a model that reaches a competitive score through longer outputs may be less suitable when users need rapid responses or when many requests must be processed.

For organizations, this makes the practical comparison broader than parameter count or index score. Local hardware costs, latency, output length, image support, data-handling needs and task accuracy all matter. No hosted pricing or measured Bolt speed is provided in the cited source, so a definitive cost comparison is not yet possible.

How Bolt Compares With Its Rivals

Bolt is not the top-scoring model in Artificial Analysis’ under-100B class. The source lists Qwen3.8 27B at 33.7, K2 Horizon MoVA 36B-A4B at 25.3 and G9v3-39A5B at 21.8, ahead of Bolt’s 21.4. The source describes Bolt as fourth overall in the class and the highest-scoring US model. That distinction may matter to organizations tracking model provenance, but it does not make Bolt the class leader.

The same comparison highlights different capabilities. The source says K2 Horizon MoVA scored higher while using fewer total and active parameters, and generated fewer tokens in the index evaluation. Bolt, however, supports images, which the source says K2 Horizon MoVA and G9v3 do not. These comparisons are based on the cited model configurations and evaluation; they do not settle which model is best for every deployment.

NVIDIA released two open-weight models on Oct. 15, with Nemotron 3.5 Super VL drawing more attention in the source account. Bolt is the first model in a new Bolt family. Its result also reflects a broader engineering question: how much capability can be delivered with fewer active parameters, and how much additional reasoning or output does the model need to reach that performance?

““Bolt is not a shrunken Ultra.””

— Thorsten Meyer, ThorstenMeyerAI.com

Speed, License and Reliability Questions

The cited evaluation does not provide measured inference speed for Bolt, and the source says no hosted API provider was listed on the model page at the time. The token-based time examples are calculations from assumed generation rates, not benchmark results. Actual response time and hardware requirements remain unknown from the material provided.

The source also does not name the permissive license shown in the model listing. That leaves a practical question for teams considering commercial or other deployments: they would need to check the model’s current license terms directly rather than rely on the broad classification alone.

Performance limits matter too. Bolt’s negative AA-Omniscience score is presented as a weakness on questions that reward correct answers and penalize confident errors, while its long-context score trails Ultra. The available material does not provide enough detail to establish how often Bolt hallucinates in ordinary use or how the benchmark results translate to specific workloads. The figures are evaluation outcomes, not guarantees of real-world performance.

What Deployment Tests Need to Show

The next useful evidence will be independent or provider-published measurements of throughput, latency and hardware use, alongside the exact model license and details of hosted availability. Those figures would let users compare the cost of running Bolt against alternatives using more than parameter counts and benchmark scores.

Teams considering self-hosting can test representative tasks, recording both answer quality and the number of generated tokens. For agent workflows, that should include completion rates and failure modes, not just a single benchmark score. For long documents or factual question answering, the cited results suggest checking accuracy directly rather than assuming the 256,000-token context window guarantees reliable reasoning across a long input.

Artificial Analysis may add provider, speed or updated evaluation information as listings change. Until then, the confirmed case for Bolt is a near-Ultra index score with far fewer parameters and a strong relative result on one agentic benchmark; the open questions are how quickly it runs, what it costs to operate, and whether its output and accuracy fit a particular workload.

Key Questions

What is Nemotron 3.5 Bolt?

Nemotron 3.5 Bolt 74B-A9B is an NVIDIA open-weight model released Oct. 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?

On Artificial Analysis Intelligence Index v4.3.2, Bolt scored 21.4 and Nemotron 3 Ultra 550B-A55B scored 22.9. That is a 1.5-point gap on that index, not evidence that the two models perform equally across all tasks.

Is Bolt the top model under 100 billion parameters?

No. The source places Bolt fourth in its under-100B class, behind Qwen3.8 27B, K2 Horizon MoVA 36B-A4B and G9v3-39A5B. It describes Bolt as the highest-scoring US model in that group.

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 when running locally, although the cited source has no measured Bolt speed; its time estimates are calculations based on assumed token rates.

Can users compare Bolt’s API price or speed now?

Not from the cited listing: the source reports no hosted API provider, price or speed measurement was available there. Users considering self-hosting would need to test performance on their own hardware and verify the model’s license terms.

Source: ThorstenMeyerAI.com