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 shipped two good open models this week. They make the US competitive again in the efficient middle of the market. They do not close the gap at the top.
On October 15, NVIDIA released two new open-weight models: Nemotron 3.5 Super VL and Nemotron 3.5 Bolt 74B-A9B. Artificial Analysis published its independent evaluation the same day.
The headline number is simple. Nemotron 3.5 Super VL scores 26 on the Artificial Analysis Intelligence Index (26.5 unrounded). Its predecessor, Nemotron 3 Super, scored 13. NVIDIA doubled its score in seven months, at the same model size, and it now needs less than half the output tokens to do the work.
Artificial Analysis calls Super VL “the leading US open weight model.” That is true among permissively licensed models. On the same day, Meta released Muse Spark Open, which scores 39 but ships under a license with commercial restrictions. And among all open-weight models, the US is still behind China and the United Arab Emirates.
So this release is two stories. NVIDIA’s own progress is excellent. America’s open-weight position has improved, but it has not changed.
What NVIDIA shipped
Both models are reasoning models, both take text and images as input, and both return text. Both are mixture-of-experts models: only a small part of the network is active for each token.
| Nemotron 3.5 Super VL | Nemotron 3.5 Bolt 74B-A9B | |
|---|---|---|
| Intelligence Index | 26 (26.5) | 21 (21.4) |
| Total / active parameters | 120B / 12B | 74B / 9B |
| Context window | 1M tokens | 256k tokens |
| Input / output | Text + image / text | Text + image / text |
| License | OpenMDW-1.1 (commercial use allowed) | Open weights; license not yet listed by AA |
| API price | $0.20 in / $0.575 out per 1M tokens | No API yet, self-host only |
| Cost per index task | $0.25 | Not measured |
Super VL is the flagship. Bolt is the first model in a new family that NVIDIA calls Bolt, built to be small enough to run on far less hardware.
One more detail matters for anyone who runs models on their own GPUs. Super VL ships in NVFP4, NVIDIA’s 4-bit number format. Artificial Analysis found the NVFP4 version scores within 0.1 points of the full-precision BF16 weights, and no single evaluation moved by more than 2.6 points. In practice, you get the full model at roughly a quarter of the memory of 16-bit weights, with almost no loss.
Infographic 01 · Four Nemotron models, seven months
Twice the score, less than half the tokens, a seventh of the price.
Nemotron 3.5 Super VL keeps the size of Nemotron 3 Super and doubles its score. Bolt gets within 1.5 points of last summer’s 550B flagship with 9B active parameters.
Intelligence Index
Higher is better. Ten evaluations.
- Nemotron 3 Super12.8
- Nemotron 3 Ultra22.9
- Nemotron 3.5 Super VL26.5
- Nemotron 3.5 Bolt21.4
Model size
Total / active parameters. Bar = total.
- Nemotron 3 Super120.6B / 12.7B
- Nemotron 3 Ultra550B / 55B
- Nemotron 3.5 Super VL120B / 12B
- Nemotron 3.5 Bolt74B / 9B
Output tokens per index task
Lower is better. Thousands of tokens.
- Nemotron 3 Super114k
- Nemotron 3 Ultra28k
- Nemotron 3.5 Super VL49k
- Nemotron 3.5 Bolt82k
Cost per index task
Lower is better. USD, median API price.
- Nemotron 3 Super$1.64
- Nemotron 3 Ultra$0.60
- Nemotron 3.5 Super VL$0.25
- Nemotron 3.5 Boltno API, not measuredn/a
Bolt has no public API yet. It can only be self-hosted, so Artificial Analysis lists no price, speed or cost per task for it. Its 82k output tokens per task is the number to watch.
Source: Artificial Analysis, Nemotron 3.5 Super and Bolt (15 Oct 2026) and model data read 16 Oct 2026. Intelligence Index v4.3.2, ten evaluations. Nemotron 3.5 Super VL output tokens per task: 49k on the model page, 46k in the launch article.
Powered by Thorsten Meyer AI
NVIDIA Nemotron 3.5 Super VL AI model
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Super VL: twice the score, less than half the tokens
The jump from Nemotron 3 Super is the cleanest result in this release. Same 120B / 12B size class, and:
- Score: 13 → 26.
- Output tokens per index task: about 114,000 → about 49,000.
- Cost per index task: $1.64 → $0.25.
That last line is the one I would underline. A model that scores twice as high and costs about 85% less per task is not a small update. It is a different product.
It also beats NVIDIA’s previous flagship. Nemotron 3 Ultra has 550B total and 55B active parameters, scores 23, and costs $0.60 per task. Super VL scores higher with less than a quarter of the parameters, at well under half the cost per task.
Speed is the other strong point. On a pre-release DeepInfra endpoint serving NVFP4 weights, Artificial Analysis measured about 630 output tokens per second with 100,000-token prompts and 1.8 seconds to the first token. That is the fastest result it has measured at that prompt length for any model scoring 20 or more. The median across the six providers now listed on the model page is lower, at 196 tokens per second, so your speed will depend on who serves it. I look at speed and price in detail in my Super VL analysis.
self-hosted AI inference GPU hardware
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Bolt: Ultra-class score on a sixth of the active parameters
Nemotron 3.5 Bolt scores 21. That is 1.5 points behind Nemotron 3 Ultra, with about a seventh of Ultra’s total parameters (74B vs 550B) and a sixth of its active parameters (9B vs 55B).
Among models under 100B total parameters, Artificial Analysis ranks Bolt fourth, and first among US models in that class. It even beats Ultra on two evaluations: AutomationBench-AA (15.5% vs 3.0%) and AA-Briefcase (912 vs 874 Elo). It trails Ultra on long-context reasoning (70% vs 79%) and Humanity’s Last Exam (25% vs 28%).
The catch is verbosity. Bolt writes about 82,000 output tokens per index task, nearly three times Ultra’s 28,000. A small model that thinks for a long time is cheap per token but not always cheap per answer. Bolt has no public API yet, so nobody can put a price on that. I cover what this means for local deployment in my Bolt analysis.
high-performance GPU for AI models
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
The US open-weight race: better, still third
Here is where the release looks less like a breakthrough.
Infographic 02 · The open-weight race
The best permissive US open model ranks 18th.
Nemotron 3.5 Super VL leads the permissively licensed US models. The top of the open-weight table still belongs to Chinese labs, and the UAE’s K2 Horizon is ahead too.
Intelligence Index, open-weight models
Each model counted once at its highest-scoring setting. Rank = position among all open-weight models Artificial Analysis lists.
The other US open-weight models
Two different “US leaders”. Artificial Analysis calls Nemotron 3.5 Super VL the leading US open-weight model. Meta’s Muse Spark Open, released the same day, scores 13 points higher but ships under the Muse Model License, which restricts commercial use.
Source: Artificial Analysis, Nemotron 3.5 Super and Bolt (15 Oct 2026) and model data read 16 Oct 2026. Intelligence Index v4.3.2, ten evaluations. Country = developer’s home country; UAE for K2 Horizon as stated by Artificial Analysis; Nex AGI not confirmed. K2 Horizon lists a $0.00 price.
Powered by Thorsten Meyer AI
Among the open-weight models Artificial Analysis currently lists, the top of the table belongs to China:
- MiMo-V2.6-Pro (Xiaomi, MIT license): 46, for $0.13 per index task.
- GLM-5.3, Step 5 and Kimi K3: 44 to 45, all under licenses with commercial conditions.
- K2 Horizon 375B A23B, from the UAE’s Institute of Foundation Models, Apache 2.0: 31.
Counting each model once, nine of the ten highest-scoring open-weight models come from Chinese labs. The tenth is Meta’s Muse Spark Open at 39. Super VL lands in 18th place. If you only count permissively licensed models, it is 11th.
That is the honest framing. Super VL leads the permissively licensed US models: it edges out Thinking Machines’ Inkling Small (25.7), passes Nemotron 3 Ultra (23) and leaves Meta’s Muse Glimmer (17) well behind. But the best permissive open model in the world scores 20 points higher, and it comes from China. Even the UAE’s best permissive model is about four points ahead.
What Muse Spark Open changes
Meta’s Muse Spark Open came out the same day. It is a far larger model (816B total, 42B active) and scores 39, thirteen points above Super VL. It also costs more to run: $0.84 per index task against Super VL’s $0.25.
The difference that matters for many buyers is the license. Super VL uses OpenMDW-1.1, which allows commercial use. Muse Spark Open uses the Muse Model License, which allows commercial use with restrictions. If your legal team needs a permissive license, Super VL is the strongest US option. If you can accept Meta’s conditions, Muse Spark Open is the more capable US model. I look at that license and the model in my Muse Spark Open analysis.
As an affiliate, we earn on qualifying purchases.
Where Nemotron 3.5 is still weak
A score of 26 is good for a 12B-active model. It is not frontier work, and the evaluation details show where the limits are.
- Terminal work: Super VL and Bolt both score 0% on Terminal-Bench 4.0, which measures real tasks in a shell. MiMo-V2.6-Pro scores 35%. If you want an open model to run coding or operations agents, Nemotron 3.5 is not it.
- Knowledge work: Super VL scores 999 Elo on AA-Briefcase, against 1,516 for MiMo-V2.6-Pro and 1,279 for Muse Spark Open.
- Hallucination: on AA-Omniscience, Super VL gets a slightly negative score (-4), which means it gives slightly more confident wrong answers than right ones. Its hallucination rate is 36%. Bolt’s is 44%.
My own experience with the previous generation points the same way. In my test of 71 model configurations on my production news pipeline, Nemotron 3 Ultra scored 2.24 out of 5, near the bottom of the field. Nemotron 3.5 has not been through that test yet. On these numbers, I would not expect it to write my news without close supervision either.
Who should use which
If you need US-origin weights under a permissive license, Nemotron 3.5 Super VL is now the default choice. It is fast, cheap at $0.25 per index task, has a 1M-token context window and reads images. It runs in NVFP4 with almost no quality loss, which matters if you own NVIDIA hardware.
If you want the most capable open model, whatever its origin, MiMo-V2.6-Pro scores 46 for $0.13 per task under an MIT license. It is also much slower (48 tokens per second). Chinese-origin weights are a policy question for many companies, not a technical one, so check before you deploy.
If you need US origin and can accept license restrictions, Muse Spark Open gives you 39 points, 13 more than Super VL, for more than three times the cost per task.
If you want a small model on limited hardware, Bolt is worth testing, but measure how many tokens it burns on your tasks before you commit.
If you run agents in a terminal, none of the US open-weight models is ready. Use a frontier API model. In my October 2026 stack, that work stays with Claude Opus 5.5 and GPT-6.1 Sol.
Verdict
NVIDIA did what a hardware company should do with its own models. It doubled the score of its mid-size model, cut its token use by more than half, made the 4-bit version almost lossless, and posted a 630-tokens-per-second result on long prompts. Bolt shows that a 9B-active model can now match June’s 550B flagship.
None of that changes the map. The strongest open-weight models still come from China, the UAE’s best is still ahead of America’s best permissive model, and the strongest US open model of the week has a restricted license. Nemotron 3.5 makes the US competitive in the efficient middle of the market. The top of the open-weight table is still somewhere else.
Sources
- Artificial Analysis: NVIDIA Releases Nemotron 3.5 Super and Nemotron 3.5 Bolt (October 15, 2026)
- Artificial Analysis: Nemotron 3.5 Super VL model page
- Artificial Analysis: Nemotron 3.5 Bolt 74B-A9B model page
- Artificial Analysis: Muse Spark Open model page
- Artificial Analysis Intelligence Index v4.3.2 data for all open-weight models, read October 16, 2026 (scores, licenses, parameters, cost and output tokens per index task)
Halloween Picks
halloween
As an affiliate, we earn on qualifying purchases.
