🔍 Read the full analysis: How Small Is Nemotron 3.5 Bolt—and What Does It Cost In Tokens? on ThorstenMeyerAI.com
Get business pricing on office and shipping supplies
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
NVIDIA released Nemotron 3.5 Bolt 74B-A9B on October 15, 2026, as an open-weight model with 74 billion total parameters and 9 billion active per token. Artificial Analysis scores it close to NVIDIA’s older 550B flagship, but ranks it fourth among models under 100B and reports about 82,000 output tokens per index task. No hosted API price or speed measurement is listed, so practical cost depends on deployment and remains uncertain.
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 much larger Nemotron 3 Ultra. But Bolt ranks fourth among models under 100 billion parameters and generated about 82,000 output tokens per index task, a volume that can add substantial time when running it locally.
Bolt uses a mixture-of-experts architecture with 74 billion total parameters and 9 billion active for each token. It accepts text and images, produces text, supports reasoning, and has a 256,000-token context window. NVIDIA released its weights for self-hosting. Artificial Analysis evaluated the reasoning version; its model page lists no hosted API provider, price, or speed measurement.
The score difference from Nemotron 3 Ultra is 1.5 points, despite Bolt having roughly one-seventh as many total parameters and one-sixth as many active parameters. Compared with Nemotron 3 Super, released in March with 120 billion total and 12 billion active parameters, Bolt’s index score is up from 12.8 to 21.4. These are benchmark results, not a promise of equivalent performance across all uses.
The benchmark profile varies by task. Artificial Analysis reports Bolt at 15.5% on AutomationBench-AA, versus 3.0% for Ultra, a test of multi-step workflows across software tools. Bolt trails Ultra on long-context reasoning, with 70% on AA-LCR against 79%, and on several knowledge-oriented measures. Both models score near zero on Terminal-Bench 4.0, so the source does not support treating Bolt as a capable shell agent.
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 Near-Flagship Score With More Waiting
Bolt’s results matter to organizations considering open-weight models they can host themselves. Its score is close to NVIDIA’s previous flagship while using fewer parameters, and Artificial Analysis ranks it first among US models under 100 billion parameters. The source also identifies image input as a distinction from some higher-ranked models in that size category. For groups with model-provenance requirements, that US ranking may be relevant, although it does not by itself establish that Bolt is the best fit for a particular deployment.
The token count complicates the efficiency story. Artificial Analysis reports about 82,000 generated tokens per index task, nearly three times Ultra’s 28,000. On a hosted service, output volume can affect billing, but no provider or price is listed for Bolt. On local hardware, generated tokens consume compute time. The source’s arithmetic estimates roughly 14 minutes for 82,000 tokens at 100 tokens per second, or 27 minutes at 50 tokens per second; those are illustrations, not measured Bolt speeds.
That trade-off may matter differently by workload. A batch process that runs unattended can tolerate long generation times more readily than an interactive assistant. Bolt’s high AutomationBench-AA score may interest developers testing agent workflows, but the benchmark does not establish that every real-world workflow will succeed or run quickly. The relevant question is not only how much a model scores, but how much time and hardware it takes to produce a useful result.
As an affiliate, we earn on qualifying purchases.
Bolt Among Models Under 100B
Artificial Analysis places Bolt fourth among models below 100 billion total parameters. The reported leaders are Qwen3.8 27B at 33.7, K2 Horizon MoVA 36B-A4B at 25.3, and G9v3-39A5B at 21.8; Bolt follows at 21.4. Its lead among US models in that group is reported as nearly four points over Muse Glimmer 30B. It is not the top-scoring model in its overall size class: Qwen3.8 is 12.3 points ahead.
The comparison also shows why parameter count alone does not settle model choice. K2 Horizon MoVA has a higher reported index score, fewer total and active parameters than Bolt, and a lower output-token count, according to the source. Bolt’s reported distinguishing features include image input and its position among US-developed models. Artificial Analysis has not measured Bolt’s speed, so any claim that its active-parameter count makes it faster on a particular system remains unverified.
The token figures come from Artificial Analysis’s index evaluation, not a fixed price list or a universal prompt. The source reports about 230 million output tokens across the evaluation, against a median of 100 million for open-weight models of Bolt’s size. It labels Bolt very verbose. That comparison describes benchmark use; individual requests can produce different output lengths.
““Those tokens are your time.””
— Thorsten Meyer, ThorstenMeyerAI.com
Price, Speed and Real-World Results
No hosted API price or provider is listed in the supplied source, and Artificial Analysis has not published a speed measurement for Bolt. As a result, its actual cost per task and runtime on specific hardware are not established. The estimates at 50 and 100 tokens per second are arithmetic examples that assume those generation rates; they are not measured performance results.
The available benchmark results also leave open how Bolt performs on particular applications, hardware configurations, and prompts. Its 256,000-token context window describes the amount of context it can accept, but the lower AA-LCR score than Ultra indicates that context capacity alone does not establish comparable reasoning over long documents. The source’s discussion of AA-Omniscience reports a negative score but is cut off before giving further explanation, so the details of that result cannot be fully assessed here.
The source classifies Bolt’s license as permissive but says the model page does not list the license name. Prospective users should check the actual model documentation for licensing terms before deployment. The supplied material also gives no independent confirmation of performance beyond the cited Artificial Analysis evaluation.
Watch for Hosting and Runtime Data
The next useful developments will be hosted-provider listings, pricing, and measured inference speeds on representative hardware. Those details would allow users to compare cost and response time directly rather than infer them from token counts or parameter totals. Independent testing could also show whether Bolt’s strong AutomationBench-AA result carries over to particular software-agent workloads.
Until those figures are available, readers can treat the Artificial Analysis scores and token counts as reported evaluation results, while treating runtime calculations as scenarios rather than benchmarks. Organizations considering the weights can test Bolt against their own tasks, hardware, and output-length requirements, and verify the license in the model documentation before using it.
Key Questions
How large is Nemotron 3.5 Bolt?
Bolt has 74 billion total parameters, with 9 billion active per token. It is a mixture-of-experts model, so the active parameter count is smaller than its total parameter count.
How many tokens does Bolt generate per task?
Artificial Analysis reports about 82,000 output tokens per Intelligence Index task. That is an evaluation average reported in the source, not a fixed output length for every user request.
Does Nemotron 3.5 Bolt have a price?
The supplied source says Artificial Analysis lists no hosted API provider or price for Bolt. Users self-hosting the open weights would need to account for their own hardware and compute costs.
Is Bolt faster than Nemotron 3 Ultra?
That is not established by the cited material. Artificial Analysis has not measured Bolt’s speed. The source’s estimates of about 14 minutes at 100 tokens per second and 27 minutes at 50 tokens per second are illustrative calculations, not Bolt benchmark results.
Is Bolt the top model under 100 billion parameters?
No. Artificial Analysis ranks Bolt fourth in that size category and first among US models in the category. The source lists Qwen3.8 27B, K2 Horizon MoVA 36B-A4B, and G9v3-39A5B ahead of it.
Source: ThorstenMeyerAI.com
Fall Picks
fall essentials
As an affiliate, we earn on qualifying purchases.
