NVIDIA has released the Nemotron ColEmbed V2 models, designed for efficient multimodal document retrieval. These models utilize a late-interaction embedding approach to improve accuracy in handling text, images, and structured visual data. They perform well on the ViDoRe V3 benchmark, making them suitable for applications like multimedia search engines and conversational AI.

The article reviews Google’s Gemini 3 Pro, highlighting its improved features over Gemini 2.5, including audio transcription capabilities and performance benchmarks compared to other AI models. It details pricing, multimodal input support, and tests involving image analysis and a city council meeting audio transcript.