Can A Model One-Seventh The Size Score At The Ultra Class? Nemotron 3.5 Bolt

🔍 Read the full analysis: Can A Model One-Seventh The Size Score At The Ultra Class? Nemotron 3.5 Bolt on ThorstenMeyerAI.com

Buying for a business?Offer from Amazon

Get business pricing on office and shipping supplies

• 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, and the model scored 21.4 on Artificial Analysis’s Intelligence Index v4.3.2, 1.5 points below the company’s 550-billion-parameter Nemotron 3 Ultra. Bolt has 74 billion total parameters, with 9 billion active per token, making the result a notable comparison in model scale—but Bolt is fourth in its under-100-billion-parameter class and generated far more output tokens during the evaluation.

Artificial Analysis lists Bolt as an open-weight, mixture-of-experts model with a 256,000-token context window, text and image inputs, text output, and reasoning capability. The evaluation cited in the report tested its reasoning version. NVIDIA also released Nemotron 3.5 Super VL on October 15; the source account says that model drew more attention, while describing Bolt as the more interesting engineering result.

The index gap between Bolt and Ultra is 1.5 points: Ultra scored 22.9 with 550 billion total parameters and 55 billion active per token. Bolt’s total parameter count is about one-seventh of Ultra’s, and its active count is about one-sixth. Against NVIDIA’s Nemotron 3 Super, released in March and scored at 12.8, Bolt scored 8.6 points higher while having 39% fewer total parameters, according to the source.

The index result does not mean Bolt matches Ultra across every task. In Artificial Analysis’s reported comparisons, Bolt scored 15.5% on AutomationBench-AA, against Ultra’s 3.0%, and had a higher AA-Briefcase Elo score. Ultra did better on long-context reasoning, knowledge-oriented tests and several other measures. Both models scored close to zero on Terminal-Bench 4.0, according to the supplied results.

A Smaller Model With Agent Strengths

Bolt’s evaluation profile may matter to teams exploring self-hosted models for agentic workflows. Its AutomationBench-AA score was five times Ultra’s in the supplied comparison, and the source says it was slightly above Nemotron 3.5 Super VL’s 14.8%. That benchmark tests multi-step work across software-as-a-service tools, so the result points to a potential strength in that particular type of task. It does not establish general agent reliability or performance in every real-world workflow.

Its size and open weights may also interest organizations that need to run models on their own infrastructure or document model provenance. Artificial Analysis ranks Bolt first among US models under 100 billion parameters, but fourth overall in that size category. The source lists Qwen3.8 27B at 33.7 and K2 Horizon MoVA 36B-A4B at 25.3, both ahead of Bolt’s 21.4. Bolt also accepts images, a capability the source says is absent from the listed K2 Horizon MoVA and G9v3 models.

However, a benchmark score is only part of the operational picture. Artificial Analysis had no hosted API provider listed for Bolt, so the source reports no price, measured speed or cost per task. The argument that nine active billion parameters could reduce compute needs relative to larger models is an architectural inference, not a measured speed result for Bolt.

How Bolt Compares With Its Peers

The Intelligence Index score places Bolt close to NVIDIA’s prior Ultra flagship, but the under-100-billion-parameter rankings put that result in perspective. Artificial Analysis’s list, counting each model once at its best configuration, gives Qwen3.8 27B a 33.7 score, K2 Horizon MoVA 36B-A4B 25.3, and G9v3-39A5B 21.8. Bolt follows at 21.4. The index therefore shows a strong result relative to NVIDIA’s own model line, not a leading score across the full size class.

The source also reports a high token count during testing: approximately 82,000 output tokens per index task. It says Bolt produced about 230 million output tokens across the index, which Artificial Analysis describes as “very verbose” compared with a 100-million-token median for open-weight models of its size. The per-task figure is reported as about 2.9 times Ultra’s 28,000 tokens and 1.5 times K2 Horizon MoVA’s 55,000.

That verbosity changes how readers should interpret the score. The source’s estimates say 82,000 tokens would take about 14 minutes at 100 tokens per second or 27 minutes at 50 tokens per second. These are arithmetic illustrations, not measured Bolt speeds; actual time depends on hardware and other factors. They show why a benchmark score per task does not, by itself, describe responsiveness or throughput.

““Bolt is one of the most verbose models in its class.””

— Thorsten Meyer, ThorstenMeyerAI.com

Speed, Cost and Reliability Remain Open

Several practical measures are not yet available in the supplied material. Artificial Analysis lists no hosted API provider for Bolt, leaving its price, measured inference speed and cost per task unknown. The source does not name the license on the model page, although it says Artificial Analysis classifies it as permissive. That makes it difficult to assess the precise reuse terms from the information provided.

The evaluation also does not establish how Bolt performs on every use case. The source reports weaker results than Ultra on long-context reasoning and several knowledge-oriented tests, including AA-LCR at 70% versus 79%. It gives Bolt an AA-Omniscience score of -19, but the supplied source text ends before explaining the full result or its interpretation. No broader conclusion about factual reliability should be drawn from that figure alone.

Finally, benchmark scores cannot settle whether Bolt will be responsive or economical on a particular local system. The token-time examples are estimates based on assumed generation rates, not measured results. Hardware, real-world task performance and deployment costs remain to be tested independently.

Provider Listings and Local Tests

The next useful evidence will be measured inference speed and pricing from a hosted provider, if one begins serving Bolt, alongside independent testing on local hardware. Such results could show whether the model’s smaller active parameter count translates into faster or cheaper use in practice, and how much its long output adds to total task time.

Teams considering Bolt can compare its benchmark strengths with their own workloads, especially multi-step tool use, image input and long-document analysis. They should treat the reported index results as historical evaluation data, not a performance guarantee, and check the actual model license and system requirements before deployment. Further clarification of the incomplete AA-Omniscience discussion would also help readers judge the knowledge-related results.

Key Questions

What is Nemotron 3.5 Bolt?

Nemotron 3.5 Bolt 74B-A9B is an NVIDIA open-weight mixture-of-experts model with 74 billion total parameters and 9 billion active per token. It accepts text and images and produces text, according to the source material.

Did Bolt match Nemotron 3 Ultra?

No. Bolt scored 21.4 on Artificial Analysis’s Intelligence Index v4.3.2, while Nemotron 3 Ultra scored 22.9. The scores are close, but the source also reports differences across individual benchmarks.

Is Bolt the highest-ranked model under 100 billion parameters?

No. Artificial Analysis ranks Bolt fourth in the under-100-billion-parameter class, while the source says it ranks first among US models in that class. Several listed models scored higher overall.

How much output did Bolt generate in the evaluation?

The source reports about 82,000 output tokens per index task and approximately 230 million output tokens across the evaluation. These figures describe benchmark output, not a fixed amount for every user prompt.

Can users access Bolt through an API?

The supplied report says Artificial Analysis listed no hosted API provider at the time. It therefore gives no hosted price, speed measurement or cost per task; open weights are available for self-hosting.

Source: ThorstenMeyerAI.com