RADIO
A vision encoder developed by NVIDIA that supports adaptive multi-scale resolutions, a dedicated video temporal embedder, and a projection layer with layer normalization, with 0.65B parameters.
Timeline (last 90 days)
Nemotron 3.5 Super VL replaces a pure attention decoder with Mamba state-space blocks, uses a self-developed RADIO vision encoder, and natively supports NVFP4 four-bit quantization on the Blackwell architecture.
According to PR 3801 in the NeMo AutoModel repository, the model's multimodal support was unified and consolidated, incorporating 512 routable experts, a RADIO vision tower, and a projection layer with normalization.
RADIO is NVIDIA's in-house vision encoder series, whose full name is Reduced-bias Adaptive Dynamic-resolution Image-to-text encOder, with the core design goal of reducing the inductive bias of traditional CLIP-like encoders.
NemotronH Omni is a multimodal reasoning model from NVIDIA that combines NemotronH's Mamba-Transformer hybrid language model with the RADIO visual encoder and an optional Parakeet audio encoder, enabling joint autoregressive reasoning over text, images, video, and audio.
RADIO is a visual foundation model proposed by NVIDIA that generates visual representations balancing semantic alignment and fine-grained features by distilling knowledge from multiple vision-language pretrained models such as CLIP, DINOv2, and SAM.
All Facts (5)
Nemotron 3.5 Super VL replaces a pure attention decoder with Mamba state-space blocks, uses a self-developed RADIO vision encoder, and natively supports NVFP4 four-bit quantization on the Blackwell architecture.
50%UnverifiedRADIO is NVIDIA's in-house vision encoder series, whose full name is Reduced-bias Adaptive Dynamic-resolution Image-to-text encOder, with the core design goal of reducing the inductive bias of traditional CLIP-like encoders.
50%UnverifiedNemotronH Omni is a multimodal reasoning model from NVIDIA that combines NemotronH's Mamba-Transformer hybrid language model with the RADIO visual encoder and an optional Parakeet audio encoder, enabling joint autoregressive reasoning over text, images, video, and audio.
50%UnverifiedRADIO is a visual foundation model proposed by NVIDIA that generates visual representations balancing semantic alignment and fine-grained features by distilling knowledge from multiple vision-language pretrained models such as CLIP, DINOv2, and SAM.
50%UnverifiedAccording to PR 3801 in the NeMo AutoModel repository, the model's multimodal support was unified and consolidated, incorporating 512 routable experts, a RADIO vision tower, and a projection layer with normalization.
50%