AIThis post was created with the assistance of artificial intelligence (AI).

🔍 Read the full analysis: What You Trade For Nemotron 3.5 Bolt’s Smaller AI Footprint on ThorstenMeyerAI.com

Before you orderOffer from Amazon

Get the latest gadgets delivered free with Prime

  • Fast, free delivery on millions of items
  • Prime Video, Amazon Music and more included
  • Member-only deals all year
Start your free Prime trial Free trial for eligible customers · Cancel anytime
As an affiliate, we earn on qualifying purchases.

TL;DR

NVIDIA released Nemotron 3.5 Bolt 74B-A9B on October 15, 2026, an open-weight model with 74 billion total parameters and 9 billion active per token. It scored 21.4 on Artificial Analysis Intelligence Index v4.3.2, close to the older 550B Nemotron 3 Ultra, but ranked fourth among models under 100B and generated about 82,000 output tokens per index task. The evaluation reports no hosted provider, speed measurement or cost per task for Bolt.

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, 1.5 points below the company’s 550B Nemotron 3 Ultra. Bolt has 74 billion total parameters and 9 billion active per token, but it ranked fourth among models under 100B and used about 82,000 output tokens per index task—an evaluation result that could mean longer waits when running it locally.

Artificial Analysis lists Bolt as a mixture-of-experts model with text and image input, text output, reasoning capability and a 256,000-token context window. NVIDIA released it with open weights for self-hosting. The organization evaluated the reasoning version. Its model page does not list a hosted API provider, so there is no reported API price, speed measurement or cost per task in the cited evaluation. The page classifies the license as permissive but does not name the license.

The index score puts Bolt close to Nemotron 3 Ultra, which scored 22.9 with 550B total and 55B active parameters. It also marks a jump from Nemotron 3 Super, a 120B-total, 12B-active model released in March that scored 12.8. Those aggregate scores do not mean the models perform alike across tasks. Artificial Analysis’ results show Bolt ahead of Ultra on AutomationBench-AA (15.5% versus 3.0%) and AA-Briefcase Elo (912 versus 876), while Ultra scored higher on long-context reasoning, knowledge-related measures and several other tests.

Artificial Analysis reports that Bolt generated about 230 million output tokens across the Intelligence Index, describing it as “very verbose” compared with a 100 million-token median for open-weight models of its size. Its per-task average was about 82,000 output tokens, compared with 28,000 for Ultra. These are evaluation statistics, not a measurement of how long Bolt will take on any particular machine.

At a glance
reportWhen: Released October 15, 2026; benchmark de…
The developmentNVIDIA released Nemotron 3.5 Bolt, a smaller open-weight model whose benchmark score approaches its previous 550B flagship, alongside a notably high output-token count.
Thorsten Meyer AI
Intelligence Index v4.3.2

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.

21.4 vs 22.9Bolt vs Nemotron 3 Ultra on the Intelligence Index. A gap of 1.5 points.
1/7Of Ultra’s total parameters (74B vs 550B), and 1/6 of its active parameters (9B vs 55B).
+8.6Points over Nemotron 3 Super (12.8), with 39% fewer total parameters.

Total parameters vs Intelligence Index

Circle size shows active parameters per token. Qwen3.8 27B (dense) for reference.

10152025303525B50B100B250B500BTotal parameters (log scale) →Intelligence Index →−1.5 points at 1/7 the sizeNemotron 3 Super · 120.6B total, 12.7B active · Intelligence Index 12.8Nemotron 3 Super · 12.8Nemotron 3 Ultra · 550B total, 55B active · Intelligence Index 22.9Nemotron 3 Ultra · 22.9Nemotron 3.5 Super VL · 120B total, 12B active · Intelligence Index 26.5Nemotron 3.5 Super VL · 26.5Nemotron 3.5 Bolt · 74B total, 9B active · Intelligence Index 21.4Nemotron 3.5 Bolt · 21.4Qwen3.8 27B · 27B total, dense · Intelligence Index 33.7Qwen3.8 27B · 33.710152025303525B100B500BTotal parameters (log scale) →Intelligence Index →−1.5 pointsNemotron 3 Super · 120.6B total, 12.7B active · Intelligence Index 12.83 Super · 12.8Nemotron 3 Ultra · 550B total, 55B active · Intelligence Index 22.93 Ultra · 22.9Nemotron 3.5 Super VL · 120B total, 12B active · Intelligence Index 26.53.5 Super VL · 26.5Nemotron 3.5 Bolt · 74B total, 9B active · Intelligence Index 21.43.5 Bolt · 21.4Qwen3.8 27B · 27B total, dense · Intelligence Index 33.7Qwen3.8 27B · 33.7

Bolt vs Ultra, evaluation by evaluation

Bars scaled within each row. Higher is better.

  1. AutomationBench-AABolt15.5%3.0%
  2. AA-BriefcaseBolt912 Elo876 Elo
  3. CritPtBolt4.3%3.1%
  4. SciCodeEven39.7%40.3%
  5. GDPval-AAUltra979 Elo1013 Elo
  6. Humanity’s Last ExamUltra24.7%28.4%
  7. AA-LCR (long context)Ultra69.7%79.3%
  8. 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

Bolt’s Agent Strength Comes With a Wait

Bolt’s most distinct reported advantage is on multi-step agentic work: AutomationBench-AA tests workflows involving SaaS tools, and Bolt’s 15.5% score was far above Ultra’s 3.0%. That result may interest developers testing open-weight models for tool-using workflows, though it does not establish that Bolt can reliably complete everyday tasks or operate a computer independently.

The token count changes how to read its near-Ultra index score. More generated tokens can require more inference time and compute when a model is self-hosted. As a simple illustration—not a measured Bolt speed—82,000 tokens would take roughly 14 minutes at 100 tokens per second or 27 minutes at 50 tokens per second. Actual performance depends on hardware, software and decoding settings, none of which Artificial Analysis supplies as a Bolt speed result. The benchmark supports a claim about score per task, not score per minute.

For organizations that need a US-developed model or image input, Bolt has features that distinguish it from some higher-ranked competitors. But its fourth-place class ranking and weaker results on long-document reasoning and knowledge tasks mean buyers should match it to a workload rather than treating the index score as a general quality guarantee.

Products Worth Considering

Bose QuietComfort Wireless Noise-Cancelling Headphones, Petal Pink

Bose QuietComfort Wireless Noise-Cancelling Headphones, Petal Pink

Bose QuietComfort Headphones (2nd Gen), Wireless Over-Ear Noise-Cancelling Headphones, Black

Bose QuietComfort Headphones (2nd Gen), Wireless Over-Ear Noise-Cancelling Headphones, Black

Logitech MX Master 4 Ergonomic Wireless Mouse with Haptic Feedback, Graphite

Logitech MX Master 4 Ergonomic Wireless Mouse with Haptic Feedback, Graphite

As an affiliate, we earn on qualifying purchases.

A Different Profile From Nemotron Ultra

Bolt is the first model in NVIDIA’s new Bolt family. The October 15 release also included Nemotron 3.5 Super VL, which the source describes as the highest-scoring permissively licensed open-weight model from the US. The focus here is Bolt’s smaller parameter count and its comparison with NVIDIA’s earlier flagship.

Among models under 100B parameters, Artificial Analysis ranked Bolt fourth overall and first among US models at the time of the cited assessment. The list was led by Qwen3.8 27B at 33.7, followed by K2 Horizon MoVA 36B-A4B at 25.3 and G9v3-39A5B at 21.8. Bolt scored 21.4. This makes it competitive within a particular national and size category, but not the leader of the overall class.

The task-level profile is mixed. Bolt outscored Ultra on AutomationBench-AA, AA-Briefcase and CritPt, but Ultra did better on AA-LCR long-context evaluation (79% to Bolt’s 70%), Humanity’s Last Exam (28% to 25%) and GDP.pdf (5.0% to 2.8%). Both models scored poorly on Terminal-Bench 4.0: Bolt recorded 0%, Ultra 0.5%. A 256,000-token context limit describes how much text the model can accept; it does not show how well it reasons across all of that material.

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

— Thorsten Meyer, ThorstenMeyerAI.com

Speed, Cost and Real-World Reliability

Artificial Analysis has not published a speed measurement or hosted price for Bolt, and its model page lists no hosted API provider. As a result, the evaluation does not show how long Bolt takes on specific hardware, what running it costs, or how its throughput compares with Ultra in practice. Token counts alone cannot settle those questions.

The reported benchmark scores also do not establish performance across every deployment. Results may differ with prompts, tools, hardware and settings. Bolt’s negative score on AA-Omniscience, which penalizes confident wrong answers, is presented in the source as a weakness, but the supplied material does not give enough detail to quantify the risk in a particular use case. The license is described as permissive, while the exact license name is not listed on the model page.

Independent Runtime Tests Will Matter

The next useful evidence will be measured inference speed and cost on clearly specified hardware, plus tests of Bolt on practical workloads. Comparisons should report output-token counts, latency and quality together; otherwise a high task score can conceal the time required to reach it. The source does not identify a date for such tests or name a hosted API launch.

Organizations considering self-hosting can evaluate Bolt against their own agent, image-input and document tasks, while checking its performance on long-context and factual-answering work. Until independent runtime results and deployment-specific tests are available, the confirmed case for Bolt rests on its published index score, parameter configuration and benchmark profile—not a demonstrated advantage in speed, price or reliability.

Key Questions

What is Nemotron 3.5 Bolt?

It 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 produces text.

How close is Bolt to Nemotron 3 Ultra?

Bolt scored 21.4 on Artificial Analysis Intelligence Index v4.3.2, compared with Ultra’s 22.9. The aggregate scores are close, but task-level results differ, including a substantial Bolt advantage on AutomationBench-AA and an Ultra advantage on long-context reasoning.

Is Bolt the top model under 100B parameters?

No. Artificial Analysis placed it fourth in the under-100B class and first among US models in that class at the time of the assessment. The listed overall leader, Qwen3.8 27B, scored 33.7 against Bolt’s 21.4.

Why does Bolt’s output-token count matter?

Artificial Analysis reports an average of about 82,000 output tokens per index task. More generated tokens may increase the time and compute required for local use, but actual latency depends on the system and settings; the cited assessment does not provide a Bolt speed measurement.

Can Bolt be used through a hosted API?

The cited Artificial Analysis model page lists no hosted API provider for Bolt. It is available as open weights for self-hosting, but the supplied material does not confirm a hosted service, price or launch date.

Source: ThorstenMeyerAI.com

You May Also Like

Meta Reportedly Made Max In Three Months. What Role Could AI Play At Home?

A report says Meta and Shanghai-based Fourier developed a humanoid prototype called Max. Meta describes the work as internal research, not a product launch.

Try Tabletop Offers Groups A Shared Space To Do More Together

Google Labs introduced Tabletop, a group collaboration experiment for shared plans and activities. U.S. adults can request access through a waitlist.

How To Create A Grok Bot Using Grok Bot On X.ai

Guide on developing a Grok Bot with xAI’s Grok system, based on recent announcement; details on process, capabilities, and next steps remain unclear.

Software Correctness & Supply Chain Security, With Sean Allen

An interview with Antithesis engineer Sean Allen examines software correctness, testing, technical debt and why supply chain attacks lack a simple fix.