NVIDIA released Nemotron 3.5 Super VL and Nemotron 3.5 Bolt 74B-A9B on October 15, 2026. Artificial Analysis measured Super VL at 26.5 on its Intelligence Index, making it the highest-scoring US open-weight model in the report; Bolt scored 21.4. Super VL’s score was nearly twice Nemotron 3 Super’s 13, while using about 60% fewer output tokens per task.

Super VL has 120B total parameters and 12B active parameters, accepts vision and text, and produces text. Its context window is 1M tokens. Bolt has 74B total and 9B active parameters, the same input and output modalities, and a 256k-token context window. Bolt is the first model in NVIDIA’s Bolt family. It ranked fourth among models below 100B parameters and first among US models in that size class.

Super VL scored ahead of NVIDIA’s Nemotron 3 Ultra at 23 and Meta’s Muse Glimmer at 17; the article gives Inkling Small’s score as both 25.7 in its key takeaways and 26 in the body. The US models trailed UAE’s K2 Horizon 375B A23B at 31 and China’s MiMo-V2.6-Pro at 46. Super VL used 46k output tokens per Intelligence Index task, compared with 114k for Nemotron 3 Super, 52k for K2 Horizon and 64k for MiMo-V2.6-Pro.

On a pre-release DeepInfra endpoint serving NVFP4 weights, Super VL generated about 630 tokens per second with 100k-token prompts, the fastest measured at that prompt length among models scoring at least 20. That was about 5x Inkling Small’s 124 tokens per second, 2.7x Nemotron 3 Ultra’s 235 and 4.5x Nemotron 3 Super’s 140. Super VL returned its first token in 1.8 seconds, versus 4.4 seconds for Ultra and 8.5 seconds for Inkling Small. In NVFP4, its index score was within 0.1 points of its BF16 score, with no component shifting by more than 2.6 points.

Bolt scored 8.6 points above Nemotron 3 Super’s 12.8 despite having 39% fewer total parameters. Against Nemotron 3 Ultra, it scored higher on AutomationBench-AA (15.5% versus 3.0%) and AA-Briefcase (912 versus 874 Elo), but lower on AA-LCR (70% versus 79%) and Humanity’s Last Exam (25% versus 28%). Bolt used 82k output tokens per task: 1.8x Super VL’s 46k and 2.9x Ultra’s 28k, but 28% fewer than Nemotron 3 Super’s 114k.

NVIDIA released Nemotron 3.5 Super VL and Nemotron 3.5 Bolt 74B-A9B on October 15, 2026. Artificial Analysis measured Super VL at 26.5 on its Intelligence Index, making it the highest-scoring US open-weight model in the report; Bolt scored 21.4. Super VL’s score was nearly twice Nemotron 3 Super’s 13, while using about 60% fewer output tokens per task.

Super VL has 120B total parameters and 12B active parameters, accepts vision and text, and produces text. Its context window is 1M tokens. Bolt has 74B total and 9B active parameters, the same input and output modalities, and a 256k-token context window. Bolt is the first model in NVIDIA’s Bolt family. It ranked fourth among models below 100B parameters and first among US models in that size class.

Super VL scored ahead of NVIDIA’s Nemotron 3 Ultra at 23 and Meta’s Muse Glimmer at 17; the article gives Inkling Small’s score as both 25.7 in its key takeaways and 26 in the body. The US models trailed UAE’s K2 Horizon 375B A23B at 31 and China’s MiMo-V2.6-Pro at 46. Super VL used 46k output tokens per Intelligence Index task, compared with 114k for Nemotron 3 Super, 52k for K2 Horizon and 64k for MiMo-V2.6-Pro.

On a pre-release DeepInfra endpoint serving NVFP4 weights, Super VL generated about 630 tokens per second with 100k-token prompts, the fastest measured at that prompt length among models scoring at least 20. That was about 5x Inkling Small’s 124 tokens per second, 2.7x Nemotron 3 Ultra’s 235 and 4.5x Nemotron 3 Super’s 140. Super VL returned its first token in 1.8 seconds, versus 4.4 seconds for Ultra and 8.5 seconds for Inkling Small. In NVFP4, its index score was within 0.1 points of its BF16 score, with no component shifting by more than 2.6 points.

Bolt scored 8.6 points above Nemotron 3 Super’s 12.8 despite having 39% fewer total parameters. Against Nemotron 3 Ultra, it scored higher on AutomationBench-AA (15.5% versus 3.0%) and AA-Briefcase (912 versus 874 Elo), but lower on AA-LCR (70% versus 79%) and Humanity’s Last Exam (25% versus 28%). Bolt used 82k output tokens per task: 1.8x Super VL’s 46k and 2.9x Ultra’s 28k, but 28% fewer than Nemotron 3 Super’s 114k.