The Compact AI Model That Makes You Pay In Tokens: Nemotron 3.5 Bolt

🔍 Read the full analysis: The Compact AI Model That Makes You Pay In Tokens: Nemotron 3.5 Bolt 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, an open-weight model that scored 21.4 on Artificial Analysis’s Intelligence Index v4.3.2, close to the company’s much larger Nemotron 3 Ultra, which scored 22.9. The result puts Bolt within 1.5 points of a 550B-parameter model, but the new model ranked fourth among models under 100B parameters and generated about 82,000 output tokens per index task.

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 inputs, text output, and reasoning capability. The model’s weights are available for self-hosting. Artificial Analysis evaluated its reasoning version.

The comparison with Nemotron 3 Ultra is notable for model size: Ultra has 550 billion total parameters and 55 billion active. Bolt’s score is 1.5 points lower on the index, despite having roughly one-seventh the total parameters and one-sixth the 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 report.

Bolt’s benchmark profile is uneven. On Artificial Analysis’s AutomationBench-AA, which tests multi-step workflows across software tools, it scored 15.5%, compared with 3.0% for Ultra. But Ultra scored higher on long-context reasoning, at 79% versus Bolt’s 70%, and on several knowledge-oriented measures. Both models scored near zero on Terminal-Bench 4.0, so the results do not support treating Bolt as a capable shell agent.

Bolt’s Agent Strength Has a Cost

The results suggest that parameter count alone does not determine performance: Bolt approaches Ultra’s overall index score with far fewer active parameters, and its AutomationBench result is substantially higher. That may make it worth testing for organizations running multi-step workflows, particularly where open weights and model provenance matter. The source report says Bolt is the highest-scoring US model under 100B parameters, a distinction that could be relevant to organizations documenting where their models come from.

But the high output-token count complicates the efficiency case. Artificial Analysis measured about 82,000 generated tokens per index task, nearly three times Ultra’s 28,000. On a hosted service, high output can affect usage costs; on local hardware, it can extend wait times. The report’s examples of roughly 14 minutes at 100 tokens per second and 27 minutes at 50 tokens per second are arithmetic illustrations, not measured Bolt speeds. Artificial Analysis had no hosted provider or speed measurement listed for the model at the time described.

For readers choosing a model, the practical question is not simply whether Bolt is close to Ultra on an index. It is whether its stronger agent-workflow result offsets its verbosity and weaker performance on some knowledge and long-document tests. That answer will depend on the task, hardware, and tolerance for latency.

How Bolt Ranks Against Rivals

Artificial Analysis placed Bolt fourth among models below 100B total parameters in the reported comparison. Qwen3.8 27B led with 33.7, followed by K2 Horizon MoVA 36B-A4B at 25.3 and G9v3-39A5B at 21.8. Bolt scored 21.4. The source report identifies it as first among US models in that size class, while emphasizing that it is not the class leader overall.

Comparisons also depend on what a model can process. Bolt accepts images as well as text; the report says K2 Horizon MoVA and G9v3 do not. Bolt’s 9B active parameters may reduce computation per generated token compared with a larger dense model, but the report cautions that this remains an architectural expectation, not a measured speed advantage. Actual throughput on particular hardware was not available.

The release included another NVIDIA model, Nemotron 3.5 Super VL, which received more attention in the source report. Bolt is the first model in a new NVIDIA Bolt family. Artificial Analysis classifies its license as permissive, but the source says the model page did not name the specific license.

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

— Thorsten Meyer, ThorstenMeyerAI.com

Speed, Cost and License Details

Several practical comparisons remain unavailable. Artificial Analysis listed no hosted API provider for Bolt, so the report gives no API price, measured speed, or cost per task. Its estimates of how long 82,000 tokens might take at different generation rates are calculations based on assumed speeds, not benchmark results for Bolt.

The reported evaluation also does not establish how Bolt will perform across real deployments or different hardware. Its index score combines results from multiple tests, while individual results point in different directions: strong agent-workflow performance, weaker long-context results than Ultra, and a negative score on AA-Omniscience, a measure that rewards correct answers and penalizes confident errors. The source excerpt does not provide enough detail to interpret that score beyond noting the weakness.

The precise license name is not shown in the source report, despite the model being classified as permissive. Further detail on licensing terms, serving options, and independent speed tests could change how organizations assess its suitability.

Deployment Tests Will Settle Trade-Offs

The next useful evidence will come from measured inference speeds and task-level costs on supported hardware, alongside the arrival of hosted API options. Those results would show whether Bolt’s smaller active parameter count produces a practical throughput advantage that offsets its high token use.

Organizations considering self-hosting can compare Bolt on their own workloads, especially agent-style software tasks, against alternatives that score higher overall or generate fewer tokens. They should also check the model’s full license terms and test knowledge accuracy and long-document reasoning before relying on it. Until independent deployment measurements are available, Bolt’s benchmark performance is confirmed, but its real-world speed and economics remain open questions.

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 and 9 billion active per token, accepts text and images, and produces text.

How close is Bolt to Nemotron 3 Ultra?

Bolt scored 21.4 on Artificial Analysis Intelligence Index v4.3.2, compared with 22.9 for Nemotron 3 Ultra. The index-score gap is 1.5 points; Bolt has substantially fewer total and active parameters.

Is Bolt the top model under 100B parameters?

No. The source report places Bolt fourth in the 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 class.

Why does Bolt’s token use matter?

Artificial Analysis reported about 82,000 output tokens per index task. More generated tokens can mean higher charges from a hosted service or longer waits on local hardware, though Bolt’s actual speed and cost have not been measured in the cited report.

Can Bolt be used as a shell agent?

The reported Terminal-Bench 4.0 results do not show useful performance: Bolt scored 0%, compared with 0.5% for Ultra. The source therefore does not support recommending it as a shell agent on that evidence.

Source: ThorstenMeyerAI.com