🔍 Read the full analysis: Nemotron 3.5 Bolt Delivers An Ultra-Class Score In Less Model Space on ThorstenMeyerAI.com
Get hardware and tech essentials delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
TL;DR
NVIDIA released Nemotron 3.5 Bolt 74B-A9B on October 15, an open-weight model with 74 billion total parameters and 9 billion active per token. It scored 21.4 on Artificial Analysis’ Intelligence Index, close to the company’s 550-billion-parameter Ultra model, but ranked fourth among models under 100 billion parameters and generated substantially more output tokens per evaluation task.
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, close to the 22.9 score of NVIDIA’s 550-billion-parameter Nemotron 3 Ultra. Bolt has 74 billion total parameters and activates 9 billion per token, but the benchmark comparison comes with a trade-off: it generated about 82,000 output tokens per index task, nearly three times Ultra’s reported average.
The score places Bolt fourth among models with fewer than 100 billion total parameters in Artificial Analysis’ ranking, and first among US models in that size class, according to the source report. The leading listed model, Alibaba’s Qwen3.8 27B, scored 33.7; K2 Horizon MoVA 36B-A4B scored 25.3, and G9v3-39A5B scored 21.8. Bolt is therefore competitive within the class, but it does not lead it.
Bolt is a mixture-of-experts model with a 262,144-token context window. It accepts text and images and produces text; the evaluated version includes reasoning. NVIDIA has made its weights available for self-hosting. Artificial Analysis listed no hosted API provider for the model at the time described in the source, leaving no listed API price, measured serving speed or cost-per-task figure.
The evaluation shows mixed strengths. Bolt scored 15.5% on AutomationBench-AA, which tests multi-step workflows across software tools, compared with 3.0% for Nemotron 3 Ultra. It also scored higher on AA-Briefcase and CritPt in the reported comparison. Ultra performed better on long-context reasoning, GDPval-AA, Humanity’s Last Exam and other knowledge-oriented measures. Both models scored 0% on Terminal-Bench 4.0 in the supplied results.
Infographic 01 · Size vs score
Bolt nearly matches NVIDIA’s old flagship at a seventh of the size.
Nemotron 3.5 Bolt 74B-A9B scores 21.4. Nemotron 3 Ultra 550B-A55B, NVIDIA’s flagship until this week, scores 22.9. The profiles differ: Bolt is stronger on agentic work, Ultra on knowledge and long documents.
Total parameters vs Intelligence Index
Circle size shows active parameters per token. Qwen3.8 27B (dense) for reference.
Bolt vs Ultra, evaluation by evaluation
Bars scaled within each row. Higher is better.
- AutomationBench-AABolt15.5%3.0%
- AA-BriefcaseBolt912 Elo876 Elo
- CritPtBolt4.3%3.1%
- SciCodeEven39.7%40.3%
- GDPval-AAUltra979 Elo1013 Elo
- Humanity’s Last ExamUltra24.7%28.4%
- AA-LCR (long context)Ultra69.7%79.3%
- GDP.pdfUltra2.8%5.0%
Not shown: AA-Omniscience, where Bolt scores −19 against Ultra’s 0 (hallucination rate 44% vs 30%), and Terminal-Bench 4.0, where both score about 0%.
Source: Artificial Analysis, Nemotron 3.5 Super and Bolt (15 Oct 2026) and Bolt model page; model data read 16 Oct 2026. Intelligence Index v4.3.2, ten evaluations.
Powered by Thorsten Meyer AI
A Smaller Model for Agent Workflows
Bolt’s result matters because its score comes from a model with roughly one-seventh of Ultra’s total parameters and one-sixth of its active parameters. That makes it a notable example of a smaller model approaching a much larger company model on an aggregate benchmark. For developers running models on their own infrastructure, active parameter count can be relevant to compute requirements, though it does not by itself establish real-world speed or hardware cost.
The more specific case for Bolt is agentic software workflows: its AutomationBench-AA result substantially exceeds Ultra’s in the figures supplied. That could make it worth evaluating for tasks involving several steps and software tools. But the benchmark does not show that Bolt is a reliable general-purpose agent, and the reported 0% Terminal-Bench score argues against treating it as a capable shell agent on that test.
Its long responses complicate the efficiency story. Artificial Analysis measured about 82,000 output tokens per index task, compared with 28,000 for Ultra. The source’s examples of 14 minutes at 100 tokens per second and 27 minutes at 50 tokens per second are arithmetic illustrations, not measured Bolt performance. Without a speed test, hardware details and task-level comparisons, the score cannot establish how quickly Bolt completes work in practice.
As an affiliate, we earn on qualifying purchases.
NVIDIA’s Model Lineup and Benchmarks
Bolt is part of a pair of open-weight models NVIDIA released on October 15. The other, Nemotron 3.5 Super VL, drew more attention in the source report. Bolt is the first model in a new family, and it follows Nemotron 3 Super, a 120-billion-parameter model with 12 billion active parameters that was released in March. The source says Super scored 12.8 on the same Intelligence Index, making Bolt’s score a substantial improvement within NVIDIA’s reported lineup.
The comparison with Ultra is not a direct measure of every capability. Artificial Analysis’ Intelligence Index combines results across evaluations, while the underlying tests show different strengths and weaknesses. Bolt’s 256K context window describes how much material it can accept, not how well it reasons across all of it: the supplied AA-LCR result was 70% for Bolt and 79% for Ultra. The model also scored −19 on AA-Omniscience, a measure that rewards correct answers and penalizes confident errors; the source characterizes this as a weakness, but does not provide enough detail here to interpret it as a general hallucination rate.
Artificial Analysis classifies Bolt’s license as permissive, but the source says the model page did not name the specific license. Organizations considering self-hosting may need to check the model’s published license and deployment terms directly rather than relying only on that broad classification.
“Bolt is not a shrunken Ultra. It is better at agentic work and weaker at knowledge and long documents.”
— Thorsten Meyer, ThorstenMeyerAI.com
As an affiliate, we earn on qualifying purchases.
Speed, Licensing and Reliability Gaps
Several practical questions remain unanswered. Artificial Analysis had no hosted API provider listed for Bolt in the source report, so there was no reported price, inference-speed measurement or cost per task. The 82,000-token figure is an evaluation average, not a promise that every prompt will use that many tokens. Actual latency will depend on the task, generation settings and hardware.
The source does not name the specific license, despite describing it as permissive. It also does not provide independent deployment tests, hardware requirements or enough detail to establish whether the benchmark advantage on AutomationBench carries over to production workflows. The reported AA-Omniscience score signals a concern on that evaluation, but does not quantify Bolt’s error rate across general use. These points remain open rather than settled by the aggregate score.
As an affiliate, we earn on qualifying purchases.
Independent Tests Will Clarify Trade-Offs
The next useful evidence will be measured inference speed and task-level token use across comparable hardware, alongside hosted API availability and pricing if a provider begins serving the model. Those results would help distinguish the model’s theoretical compute profile from the time users actually wait.
Developers and organizations can also compare Bolt with Ultra and other sub-100-billion-parameter models on their own agent workflows, long-document tasks and accuracy requirements. Before deployment, they should verify the model’s specific license and assess outputs for reliability. Until those details are available, the benchmark supports a narrow conclusion: Bolt combines a score near NVIDIA’s larger Ultra with a reported advantage on one agent-workflow test, while its ranking, verbosity and weaker results on several other evaluations limit broader claims.
As an affiliate, we earn on qualifying purchases.
Key Questions
What is Nemotron 3.5 Bolt?
It is an open-weight NVIDIA model with 74 billion total parameters and 9 billion active per token. It accepts text and images, produces text and has a reported 262,144-token context window.
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 aggregate scores are close, but individual evaluations show different strengths.
Is Bolt the highest-ranked model under 100 billion parameters?
No. The source report places it fourth in its size class. It lists Qwen3.8 27B, K2 Horizon MoVA 36B-A4B and G9v3-39A5B ahead of it.
What is the main trade-off reported for Bolt?
Bolt performed strongly on AutomationBench-AA but generated about 82,000 output tokens per Intelligence Index task, compared with 28,000 for Ultra. Actual completion speed has not been established in the source material.
Can Bolt be used through a hosted API?
The source report says Artificial Analysis listed no hosted API provider at the time. It does not establish whether providers have added access since that report was published.
Source: ThorstenMeyerAI.com
Fall Picks
fall essentials
As an affiliate, we earn on qualifying purchases.
