NVIDIA releases Nemotron 3.5 Super VL and Bolt weights, tops US open-weight models with a third-party score of 26 — but its "beats Opus 5" claim applies only after post-training, on specialized tasks

NVIDIA released the weights of its multimodal model Nemotron 3.5 Super VL on October 15. On the same day, it also released a smaller model, Bolt 74B-A9B. Independent evaluator Artificial Analysis gave Super VL a score of 26 on its Intelligence Index, the highest of any US open-weight model at the time of release. The model is also fast, generating about 630 tokens per second.

However, the number behind the headlines, that it "beat Claude Opus 5 and GPT-6.1 Sol," comes with conditions. It only covers five specialized tasks, measured after companies post-trained the model on their own data. This article looks separately at the model's out-of-the-box ability as measured by a third party and at how NVIDIA has presented it.

What was released

According to NVIDIA's official blog and its developer blog (both primary sources), Nemotron 3.5 Super VL has 120B total parameters and 12B active parameters. It uses a hybrid Mamba-Transformer architecture, can read images and handles a context of up to 1 million tokens.

The weights of Bolt 74B-A9B were released on the same day. An even smaller model, Bolt E4B, is scheduled for release on October 23. NVIDIA's announcement for the family focuses on "powering specialized agents," not on taking the top spot in general-purpose chat. That is worth keeping in mind from the start.

A third-party score of 26, the top US open-weight model at release

Artificial Analysis published an evaluation article on the day of release. It wrote:

This score was measured by a third party, not reported by NVIDIA itself. The previous Nemotron 3 Super scored 13, so the score has doubled. The new model also beat the previous flagship, Nemotron 3 Ultra (23). At 120B total parameters, it is clearly ahead of US models of similar size.

But its position as the top US model lasted only one day. On October 16, Artificial Analysis gave Meta's Muse Spark Open a score of 39 and Reflection AI's Beam a score of 30. Muse Spark Open became the top US open-weight model, and Nemotron fell to third. Still, Muse Spark Open has 816B total and 42B active parameters, and Beam has 501B total and 23B active parameters. Both are more than four times the size of Nemotron, so they are not a like-for-like comparison. Nemotron's strengths are its small size and speed.

Outside the US, the top Chinese open-weight models are MiMo-V2.6-Pro, with a score of 46, and DeepSeek V4.1-Flash-1017, with 45. Among closed models, Claude Opus 5 scores 51, GPT-6.1 Sol (max) scores 52 and Opus 5.5 scores 58. On general ability, Nemotron 3.5 Super VL scores about half as high as frontier models.

About 630 tokens per second

According to the Artificial Analysis article, the model generates about 630 tokens per second. The same evaluator measured Muse Spark Open at 118 tokens per second, so Nemotron is roughly five times faster. The small 12B active parameter count and the hybrid Mamba design appear intended to deliver both speed and long context.

Agent workloads often call a model dozens of times to complete a single task. As a result, the speed and cost of each call often matter more to the total cost than the quality of any single answer. Under the "AI tokenomics" approach, where companies pick a cheap, fast model for each task, this speed may matter more than the benchmark score.

What lies behind "beats Opus 5"

A VentureBeat article (a secondary source) ran the headline that "post-trained Nemotron 3.5 Super VL beats Claude Opus 5 and GPT-6.1 Sol on specialized task benchmarks." Three points stand out.

First, the comparison used the model after post-training. The figures were measured after companies retrained the model on their own data. They are not scores for the base Nemotron as downloaded. The base model's general ability is the score of 26 discussed above.

Second, the comparison covered only five specialized tasks. A model trained for a specific narrow job often beats a larger general-purpose model at that job. The result does not show that the model is "smarter than Opus 5."

Third, the comparison comes from NVIDIA's own announcement. No third party has yet rerun these five tasks under the same conditions.

That said, the figures do matter in practice. The weights of Opus 5 and GPT-6.1 Sol are not public, so companies cannot post-train those models themselves. With an open-weight model, a company can retrain it on its own data and bring it up to the level of larger models on narrow tasks. The most accurate reading is that NVIDIA's announcement shows this use case.

What it means

This announcement needs to be read as two separate numbers. The general score of 26, measured by a third party, is strong for a US model of its size, but about half as high as frontier models. The "beats Opus 5" claim, by contrast, comes from just five tasks, after companies retrained the model on their own data. NVIDIA is selling a way to build cheap, fast agents for narrow tasks by combining about 630 tokens per second with weights companies can retrain themselves. It has not taken the top spot on general intelligence.

The next things to watch are Bolt E4B, scheduled for release on October 23, and whether a third party reruns the comparison on these five specialized tasks.