Nvidia and DeepSeek are making aggressive moves in multimodal AI, unveiling new agent platforms that integrate vision, audio, and language processing.
Nvidia's Nemotron 3 Nano Omni combines vision, audio, and speech capabilities into a single model, enabling real-time multimodal interactions. Meanwhile, DeepSeek's V4 model features 1.6 trillion parameters in a Mixture-of-Experts architecture, a 384K context window, and tenfold performance improvements over its predecessor.
Other developments include Qwen 3.6, which surprisingly outperforms larger models despite having only 27 billion parameters, and ongoing quality concerns with Anthropic's Claude Code tool, which has been plagued by reports of errors and poor support.
The video also cautions that running large language models locally does not guarantee privacy, highlighting potential data leaks in local deployments.