By Thorsten Meyer | October 16, 2026 | ThorstenMeyerAI.com
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
NVIDIA’s new small model nearly matches its old flagship. It is not the best model in its class, and it talks a lot before it answers.
On October 15, 2026, NVIDIA released two open-weight models. Most of the attention went to Nemotron 3.5 Super VL, now the highest-scoring permissively licensed open-weight model from the US. The second one is the more interesting engineering result: Nemotron 3.5 Bolt 74B-A9B, the first model in a new Bolt family.
Bolt scores 21.4 on the Artificial Analysis Intelligence Index v4.3.2. NVIDIA’s previous flagship, Nemotron 3 Ultra 550B-A55B, scores 22.9. That is a gap of 1.5 points, from a model with about a seventh of the total parameters (74B vs 550B) and a sixth of the active parameters (9B vs 55B).
That is the headline. Two things complicate it. Bolt is fourth in its own size class, not first. And it uses 82,000 output tokens per index task, nearly three times what Ultra needs. If you run it on your own hardware, those tokens are your time.
What NVIDIA shipped
- Architecture: Mixture of Experts, 74 billion total parameters, 9 billion active per token.
- Context window: 256k tokens (262,144).
- Input: text and images. Output: text.
- Reasoning: yes. Artificial Analysis evaluated the reasoning version.
- Availability: open weights for self-hosting. Artificial Analysis lists no hosted API provider yet, so it has no price, no speed measurement and no cost per task. Its license is classed as permissive, but the license name is not listed on the model page.
Against NVIDIA’s own line-up, the progress is steep. Nemotron 3 Super (120B total, 12B active), released in March, scored 12.8. Bolt adds 8.6 points with 39% fewer total parameters.
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
As an affiliate, we earn on qualifying purchases.
Where Bolt beats Ultra, and where it doesn’t
An index score close to Ultra’s could hide a very different profile, and in this case it does. Bolt is not a shrunken Ultra. It is better at agentic work and weaker at knowledge and long documents.
| Evaluation | Bolt 74B-A9B | Nemotron 3 Ultra | Edge |
|---|---|---|---|
| AutomationBench-AA | 15.5% | 3.0% | Bolt, by a wide margin |
| AA-Briefcase (Elo) | 912 | 876 | Bolt |
| CritPt | 4.3% | 3.1% | Bolt, both very low |
| GDPval-AA (Elo) | 979 | 1013 | Ultra |
| SciCode | 39.7% | 40.3% | Even |
| AA-LCR (long context) | 70% | 79% | Ultra |
| Humanity’s Last Exam | 25% | 28% | Ultra |
| GDP.pdf | 2.8% | 5.0% | Ultra |
| AA-Omniscience | −19 | 0 | Ultra, clearly |
| Terminal-Bench 4.0 | 0% | 0.5% | Neither |
The AutomationBench result is the one that matters for agents. It measures multi-step workflows across SaaS tools. Bolt at 15.5% is five times Ultra’s score and slightly above the new Super VL (14.8%). For a 9B-active model, that is a real result.
The other end of the table matters just as much. Neither model can do anything useful on Terminal-Bench 4.0, so do not plan on Bolt as a shell agent. And AA-LCR, which tests reasoning over long documents, is nine points worse than Ultra. A 256k context window is only useful if the model can reason over what is in it.
high performance GPU server for AI
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Fourth in its class, first from the US
Artificial Analysis places Bolt fourth among models under 100B total parameters, and first among US models in that class. Here is the class as it stands, counting each model once at its best configuration:
- Qwen3.8 27B (Alibaba, Apache 2.0): 33.7
- K2 Horizon MoVA 36B-A4B (Institute of Foundation Models, UAE, Apache 2.0): 25.3
- G9v3-39A5B (AI9Stars, Apache 2.0): 21.8
- Nemotron 3.5 Bolt 74B-A9B (NVIDIA): 21.4
- K2 Horizon 7B (Institute of Foundation Models): 20.6
- Qwen3.6 35B-A3B (Alibaba): 18.2
- Muse Glimmer 30B (Meta): 17.5
- Gemma 4 26B-A4B (Google): 16.7
Qwen3.8 27B is 12 points ahead. That is not a close race. K2 Horizon MoVA scores four points more with half of Bolt’s total parameters and fewer than half its active parameters, and it uses fewer tokens.
So why care about Bolt? Three reasons. It is the strongest US model under 100B by almost four points over Muse Glimmer, which matters to organizations that must document model provenance. It accepts images, which K2 Horizon MoVA and G9v3 do not. And with 9B active parameters, each generated token needs far less compute than a 27B dense model, which is an architectural advantage on your own hardware. Artificial Analysis has not measured Bolt’s speed, so treat that last point as theory until someone does.
Infographic 02 · The small-model class
Fourth in its class, first from the US, and the most talkative.
Among open-weight models under 100B total parameters, Bolt ranks fourth on the Intelligence Index. Of the models here with reported figures, it also writes the most output tokens per task. With no hosted API yet, those tokens cost time on your own hardware.
Under-100B open-weight models: score and verbosity
Best configuration per model. Parameters shown as total / active.
- Qwen3.8 27BAlibaba · China · 27B denseIndex33.7Tokens/task67k
- K2 Horizon MoVA 36B-A4BIFM · UAE · 36B / 4BIndex25.3Tokens/task55k
- G9v3-39A5BAI9Stars · 39B / 5BIndex21.8Tokens/tasknot reported
- Nemotron 3.5 Bolt 74B-A9BNVIDIA · US · 74B / 9BIndex21.4Tokens/task82k
- K2 Horizon 7BIFM · UAE · 7B denseIndex20.6Tokens/task72k
- Qwen3.6 35B-A3BAlibaba · China · 35B / 3BIndex18.2Tokens/task35k
- Muse GlimmerMeta · US · 30BIndex17.5Tokens/task14k
- Gemma 4 26B-A4BGoogle · US · 25B / 3.8BIndex16.7Tokens/tasknot reported
- Nemotron 3.5 LightningNVIDIA · US · 32B / 3.6BIndex12.9Tokens/task37k
What 82k tokens mean in time
Arithmetic only: tokens per task ÷ decode speed. Bolt’s speed has not been measured.
| Decode speed | Bolt (82k) | Ultra (28k) |
|---|---|---|
| 200 t/s | 7 min | 2.4 min |
| 100 t/s | 14 min | 4.7 min |
| 50 t/s | 27 min | 9.5 min |
Weight memory, estimated
Parameters × bytes per weight. Weights only, before KV cache and serving overhead.
- BF16~148 GB
- 8-bit~74 GB
- 4-bit~37 GB
For scale: Nemotron 3 Ultra at 4-bit is about 275 GB. No hosted API for Bolt is listed yet.
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
large context window AI processing device
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
The catch: 82,000 tokens per task
Bolt is one of the most verbose models in its class. It generated about 230 million output tokens to complete the Intelligence Index, which Artificial Analysis rates as “very verbose” against a median of 100 million for open-weight models of its size. Per task, that is 82,000 output tokens:
- 2.9× Nemotron 3 Ultra (28k)
- About 1.7× Nemotron 3.5 Super VL (49k on the current model page; NVIDIA’s launch comparison quoted 46k and 1.8×)
- 1.2× Qwen3.8 27B (67k)
- 1.5× K2 Horizon MoVA (55k)
- 28% fewer than Nemotron 3 Super (114k)
On a hosted API, verbose models cost money. On your own GPU, they cost time, and time is the budget you actually have. Simple arithmetic shows the problem. These are not measurements. At 100 tokens per second, 82,000 tokens take about 14 minutes. At 50 tokens per second, about 27 minutes. Ultra’s 28,000 tokens would take under 5 minutes at 100 tokens per second.
Some of this is the price of reasoning. Small models often compensate for limited capacity by thinking longer. But it means Bolt’s “Ultra-class score” is an Ultra-class score per task, not per minute. For batch jobs that run overnight, that is fine. For anything a person waits for, it is not.
I have seen the same pattern in my own work. In my news-pipeline test of 71 model configurations, the best local model, Qwen3.8-Flash-Next, reached its best score only with thinking enabled, and then it needed about 312 seconds per article. Local quality is often available. Local speed at that quality is the hard part.
AI training and inference hardware
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Knowledge and hallucination: the weak spot
AA-Omniscience rewards correct answers and penalizes confident wrong ones. Bolt scores −19, meaning it gives clearly more wrong answers than right ones. Its hallucination rate is 44%, against 30% for Ultra and 36% for Super VL.
This is common for small models. Knowledge takes parameters, and 74B total is not much. The practical rule is simple: do not use Bolt as a source of facts. Give it the facts through retrieval or the prompt, and check what it writes. The 2.8% on GDP.pdf points in the same direction. Professional document reasoning with strict all-pass grading is beyond it.
What it takes to run it
Because there is no hosted API, the useful question is hardware. Here is the weight-memory arithmetic. These are simple estimates for the weights alone, before KV cache, activations and serving overhead, not measured requirements:
- BF16 (16-bit): about 148 GB
- 8-bit: about 74 GB
- 4-bit: about 37 GB
For comparison, Nemotron 3 Ultra needs roughly 275 GB at 4-bit, and Super VL roughly 60 GB. At 4-bit, Bolt’s weights fit on a single high-memory workstation GPU or a well-equipped Apple Silicon machine, which Ultra’s do not. That is the real meaning of “Ultra-class at a seventh of the size”: it moves this score into the range of hardware a small team can own.
One caution. NVIDIA reported near-lossless NVFP4 results for Super VL, within 0.1 points of BF16. That result is for Super VL. Nobody has published the equivalent for Bolt yet, so measure quality after you quantize. I wrote about why 4-bit models can behave differently from their full-precision versions; the effect can go in either direction.
Who should use which
If you need the best model under 100B and have no origin constraint, use Qwen3.8 27B. It scores 12 points more, it is Apache 2.0 licensed, and it is served by 12 providers if you want to test before you host. It is slow on those APIs (about 41 tokens per second) and not cheap at $1.01 per index task, but no small model beats its score.
If you need a US-origin model you can run yourself, Bolt is now the strongest option under 100B. It is clearly better than Muse Glimmer (17.5) and Gemma 4 26B-A4B (16.7, Google). Plan for long outputs, keep a retrieval step for facts, and measure throughput on your hardware before you commit.
If you run agent workflows across business tools, Bolt’s AutomationBench result is worth a test. 15.5% beats Ultra and Super VL. It is still low in absolute terms, so test on your own workflows.
If you can use a hosted API, look at Super VL first. It scores five points more, costs $0.25 per index task across six providers, and uses fewer tokens than Bolt. I cover it in a separate analysis, and the whole release in the overview.
If you need knowledge, long documents or a shell agent, Bolt is the wrong tool. Its scores on AA-Omniscience, AA-LCR and Terminal-Bench 4.0 say so plainly.
None of this changes my own setup. In my October 2026 AI stack, Opus 5.5 builds and GPT-6.1 Sol digs. A 21-point model does not change that. It answers a different question: what can you run when the data must not leave your building?
Verdict
Nemotron 3.5 Bolt is a real step for NVIDIA. It takes the score of a 550B flagship and puts it into a 74B model whose 4-bit weights fit on hardware a small team can own. It also beats that flagship on agentic automation, by a wide margin.
It is not the best small model. Qwen3.8 27B leads the class by 12 points, and K2 Horizon MoVA scores higher with half the size. Bolt also pays for its score in tokens. 82,000 per task is a lot of waiting on your own GPU, and a 44% hallucination rate means it needs facts supplied, not recalled.
Bolt is worth testing if you need a US-origin model that runs locally and accepts images. If you don’t, the class leader comes from Alibaba.
Sources
- Artificial Analysis: NVIDIA Releases Nemotron 3.5 Super and Nemotron 3.5 Bolt (October 15, 2026)
- Artificial Analysis: Nemotron 3.5 Bolt 74B-A9B model page
- Artificial Analysis: Nemotron 3.5 Super VL model page
- Artificial Analysis Intelligence Index v4.3.2 data for all compared models, read October 16, 2026 (scores, evaluation results, output tokens per task, parameters)
Halloween Picks
halloween
As an affiliate, we earn on qualifying purchases.
