NVIDIA Releases Nemotron 3.5 Super and Nemotron 3.5 Bolt

NVIDIA Releases Nemotron 3.5 Super and Nemotron 3.5 Bolt

NVIDIA has released Nemotron 3.5 Super VL and Nemotron 3.5 Bolt 74B-A9B. Nemotron 3.5 Super VL scores 26.5 on the Artificial Analysis Intelligence Index, making it the highest-scoring US open-weight model in the evaluation. It achieves roughly twice the score of its predecessor, Nemotron 3 Super, while using about 60% fewer output tokens per task.

Nemotron 3.5 Bolt 74B-A9B is the first model in NVIDIA's new Bolt family. It scores 21.4 on the Intelligence Index.

Key results

• Leading US open-weight model: Nemotron 3.5 Super VL scores 26.5, ahead of Thinking Machines' Inkling Small at 25.7, NVIDIA's Nemotron 3 Ultra at 23, and Meta's Muse Glimmer at 17.
• Bolt approaches Ultra's score: Nemotron 3.5 Bolt scores 21.4, compared with 23 for Nemotron 3 Ultra 550B A55B. Bolt has 74 billion total parameters and 9 billion active parameters, versus 550 billion and 55 billion, respectively, for Ultra.
• Minimal NVFP4 score reduction: Nemotron 3.5 Super VL's NVFP4 release scores within 0.1 points of its BF16 version on the Intelligence Index. No individual component changes by more than 2.6 points.

Model specifications

Nemotron 3.5 Super VL

• Context window: 1 million tokens
• Inputs: Vision and text
• Output: Text
• Parameters: 120B total, 12B active

Nemotron 3.5 Bolt 74B-A9B

• Context window: 256,000 tokens
• Inputs: Vision and text
• Output: Text
• Parameters: 74B total, 9B active

Nemotron 3.5 Super: US open-weight leader

Nemotron 3.5 Super VL is the highest-scoring open-weight model from the United States in the Intelligence Index results, with a score of 26.5. It narrowly exceeds Thinking Machines' Inkling Small, which scores 25.7, and remains ahead of NVIDIA's Nemotron 3 Ultra at 23 and Meta's Muse Glimmer at 17.

The United States remains behind other regions in the open-weight rankings. The United Arab Emirates' K2 Horizon 375B A23B scores 31, while China's MiMo-V2.6-Pro scores 46.

Token usage

Nemotron 3.5 Super VL generates an average of 46,000 output tokens per Intelligence Index task. That is approximately 60% fewer than Nemotron 3 Super, which uses 114,000 tokens. Super VL also uses fewer tokens than K2 Horizon 375B A23B at 52,000 and MiMo-V2.6-Pro at 64,000.

Although its token efficiency has improved, Nemotron 3.5 Super VL remains behind the Intelligence Index versus output-tokens-per-task frontier.

Output speed

On a pre-release @DeepInfra endpoint serving NVFP4 weights, Nemotron 3.5 Super VL reaches approximately 630 tokens per second with 100,000-token prompts. Among models scoring at least 20 on the Intelligence Index, this was the highest measured output speed at that prompt length.

At the same prompt length, Super VL is approximately five times faster than Inkling Small at 124 tokens per second, 2.7 times faster than Nemotron 3 Ultra at 235 tokens per second, and 4.5 times faster than Nemotron 3 Super at 140 tokens per second.

Its time to first token is 1.8 seconds, compared with 4.4 seconds for Nemotron 3 Ultra and 8.5 seconds for Inkling Small.

Nemotron 3.5 Bolt

Nemotron 3.5 Bolt 74B-A9B is the first model in NVIDIA's Bolt family. With a score of 21.4 on the Intelligence Index, it ranks fourth among models with fewer than 100 billion parameters and first among US models in that group.

Bolt scores 8.6 points higher than Nemotron 3 Super, which scores 12.8, while using 39% fewer total parameters. Compared with Nemotron 3 Ultra, Bolt performs better on AutomationBench-AA, scoring 15.5% versus 3.0%, and on AA-Briefcase, scoring 912 versus 874 Elo.

Nemotron 3.5 Bolt trails Nemotron 3 Ultra on AA-LCR, at 70% versus 79%, and on Humanity's Last Exam, at 25% versus 28%.

Bolt token usage

Nemotron 3.5 Bolt uses 82,000 output tokens per Intelligence Index task. That is 1.8 times the 46,000 tokens used by Nemotron 3.5 Super VL and 2.9 times the 28,000 used by Nemotron 3 Ultra. Bolt still uses 28% fewer output tokens than Nemotron 3 Super, which uses 114,000 tokens per task.