Nemotron-3.5-Super-VL-120B-A12B
An open-source vision-language model family released by NVIDIA. It uses a Mixture-of-Experts architecture with 120B total parameters and 12B activated parameters, supports multimodal reasoning, and is offered in base, NVFP4, and BF16 versions along with a LiLiCorr speculative decoding scheme.
Timeline (last 90 days)
NVIDIA released the Nemotron-3.5-Super-VL-120B-A12B model family, covering a base version as well as multiple quantization formats including NVFP4 and BF16, and introduced the LiLiCorr speculative decoding scheme.
In the name Nemotron-3.5-Super-VL-120B-A12B, '120B-A12B' most likely indicates a MoE architecture with 120B total parameters and 12B activated parameters, and 'VL' points to vision-language multimodal capability.
In the 120B-A12B model, the total model weights are 120 billion parameters, but only about 12 billion parameters actually participate in computation during each token's forward pass.
All Facts (3)
In the name Nemotron-3.5-Super-VL-120B-A12B, '120B-A12B' most likely indicates a MoE architecture with 120B total parameters and 12B activated parameters, and 'VL' points to vision-language multimodal capability.
50%UnverifiedIn the 120B-A12B model, the total model weights are 120 billion parameters, but only about 12 billion parameters actually participate in computation during each token's forward pass.
50%UnverifiedNVIDIA released the Nemotron-3.5-Super-VL-120B-A12B model family, covering a base version as well as multiple quantization formats including NVFP4 and BF16, and introduced the LiLiCorr speculative decoding scheme.
50%