Size, Scores, And Tokens: Understanding Nemotron 3.5 Bolt

🔍 Read the full analysis: Size, Scores, And Tokens: Understanding Nemotron 3.5 Bolt on ThorstenMeyerAI.com

Before you orderOffer from Amazon

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

Start your free Prime trial →Free trial for eligible customers · Cancel anytime

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, 1.5 points below the company’s 550B-parameter Nemotron 3 Ultra. Bolt has 74 billion total parameters and activates 9 billion per token, but it ranked fourth among models under 100 billion parameters and generated about 82,000 output tokens per index task, a factor that can affect response time and compute use.

Bolt is a mixture-of-experts model with a 256,000-token context window. It accepts text and images and produces text; the evaluated version supports reasoning. NVIDIA has released open weights for self-hosting. Artificial Analysis lists no hosted API provider for Bolt, so the source material gives no measured API price, inference speed or cost per task. It also says the model page classifies the license as permissive but does not name the license.

The Intelligence Index score places Bolt close to Nemotron 3 Ultra 550B-A55B, which scored 22.9. The comparison is not a direct measure of speed or operating cost: Artificial Analysis reports that Bolt generated about 230 million output tokens across the index, or roughly 82,000 per task. The source describes this as very verbose compared with a 100 million-token median for open-weight models of similar size.

Bolt showed a mixed performance profile against Ultra. On AutomationBench-AA, which tests multi-step workflows across software tools, Bolt scored 15.5% versus Ultra’s 3.0%. It also led on AA-Briefcase and CritPt in the supplied comparison. Ultra scored higher on long-context reasoning, knowledge-related measures and several other tests, including AA-LCR (79% to Bolt’s 70%) and Humanity’s Last Exam (28% to 25%). Both models scored 0% on Terminal-Bench 4.0 in the comparison.

Bolt’s Agent Strength Comes With a Time Cost

The result highlights a trade-off for teams choosing open-weight models: a strong aggregate score does not mean strength on every task, nor does it show how quickly a model completes work. Bolt’s AutomationBench result could interest developers building agents that use business software, while its weaker long-document result and lack of performance on Terminal-Bench limit what the reported scores support.

Output volume is also relevant for practical deployment. The source gives arithmetic examples, not measured Bolt speeds: at 100 tokens per second, 82,000 tokens would take about 14 minutes to generate; at 50 tokens per second, about 27 minutes. Actual completion time depends on hardware, configuration and other factors. For self-hosted users, long outputs can consume compute capacity and increase waiting time. Without measured speed or a hosted price, Bolt’s efficiency cannot be established from parameter counts alone.

Its standing is strong within one particular slice of the market, but it is not the class leader. Artificial Analysis ranks it fourth under 100B total parameters and first among US models in that group. The highest-listed model, Qwen3.8 27B, scored 33.7; K2 Horizon MoVA 36B-A4B scored 25.3. The distinction may matter to organizations tracking model origin, while image input may be useful for tasks that competing models without image support cannot handle. Those factors do not make Bolt the highest-scoring model overall in its size class.

From Super to a Smaller Bolt Model

NVIDIA announced two open-weight models on October 15: Nemotron 3.5 Super VL and Bolt, the first model in a new Bolt family. The supplied report says Super VL drew more attention, while Bolt stood out as an engineering comparison with NVIDIA’s older flagship. It also describes Super VL as the highest-scoring permissively licensed open-weight model from the US; that statement concerns Super VL, not Bolt.

Bolt follows Nemotron 3 Super, a 120B-total, 12B-active model released in March that scored 12.8 on the same index, according to the source. Bolt’s 21.4 is 8.6 points higher, despite having fewer total parameters. Scores from one benchmark suite, however, do not by themselves establish broad superiority or lower operating cost.

The index comparison also should not be confused with a universal ranking. Artificial Analysis places Bolt fourth among models below 100B parameters, and the source’s table lists models from several developers with different architectures and licenses. Bolt’s 256K context window describes the amount of input it can accept; it does not guarantee reliable reasoning across every long document.

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

— Thorsten Meyer, ThorstenMeyerAI.com

Speed, Cost and Knowledge Limits

Artificial Analysis has not published a speed measurement for Bolt in the material provided, and there is no listed hosted API provider or price. The token counts are evaluation observations; estimated completion times based on assumed token rates are calculations, not benchmark measurements. Hardware, software and generation settings could change real-world results.

The source is also incomplete about Bolt’s knowledge and hallucination profile. It reports a score of −19 on AA-Omniscience, a measure that rewards correct answers and penalizes confident wrong ones, but the supplied text ends before explaining the result or giving more detail. The score should not be treated as a standalone measure of hallucination frequency. The listed benchmark results also do not establish how Bolt performs across all real-world agent, coding or document tasks.

The model page’s license is described as permissive, but its exact license name is not provided in the source. Prospective users would need to check NVIDIA’s model documentation and terms before deployment. The source does not provide information on independent replication of the results or on performance under specific hardware configurations.

Wait for Deployment Measurements

The next useful evidence will be measured inference speed, hardware requirements and repeatable task-level testing. Those results would help determine whether Bolt’s lower active-parameter count translates into a practical advantage on local systems, and how its long outputs affect latency and throughput.

Teams evaluating the model can compare its performance on their own workflows, including whether its agentic strengths outweigh weaker scores on long-context and knowledge tests. They should also verify the current license terms and check whether a hosted provider becomes available. Until such information is published, the confirmed picture is a strong index score paired with high output volume and a mixed benchmark profile.

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 activates 9 billion per token.

How close is Bolt to Nemotron 3 Ultra?

Bolt scored 21.4 on Artificial Analysis’s Intelligence Index v4.3.2, compared with 22.9 for Nemotron 3 Ultra 550B-A55B. That is a 1.5-point score difference, not a comparison of speed or cost.

Is Bolt the top model under 100 billion parameters?

No. Artificial Analysis ranks Bolt fourth in the under-100B group in the supplied report, although it ranks first among US models in that group. Qwen3.8 27B and K2 Horizon MoVA 36B-A4B scored higher.

Why might Bolt take longer to answer?

The evaluation reports about 82,000 output tokens per index task. Longer output can take more time and compute to generate, but no measured Bolt speed is available in the cited material, so actual response times will vary by hardware and setup.

Can Bolt be used through a hosted API?

The source says Artificial Analysis listed no hosted API provider for Bolt at the time of its report. NVIDIA has released open weights for self-hosting; availability of hosted access may change.

Source: ThorstenMeyerAI.com