AI Radar

Noteworthy AI conversations and writing, summarized from publisher notes.

NVIDIA's Nemotron Family and the Rise of Reasoning Models

· Practical AI

Joey Conway, Senior Director of Product Management for AI Models at NVIDIA, joins the Practical AI podcast to discuss NVIDIA's recently announced Nemotron family of open-weight models. Conway frames reasoning as the latest major inflection point in language model capabilities—following BERT and ChatGPT—crediting DeepSeek for pioneering open-weight reasoning models. He explains that reasoning models can decompose complex queries into sub-problems and walk through solution steps, much like showing work on a math problem, unlocking capabilities in math, science, coding, and tool calling that previous models couldn't achieve. The Nemotron family consists of three Llama-based variants: Nano (8B parameters, targeting ~24GB GPUs), Super (49B, fitting on a single H100/A100), and Ultra (253B, fitting in roughly half a node). NVIDIA used neural architecture search to shrink the model sizes while maintaining accuracy, and uniquely packaged both reasoning and non-reasoning capabilities into a single model to ease enterprise deployment. All training data, techniques, and software (via the NeMo Framework and NeMo RL) are open-sourced. Conway also briefly covers Parakeet, NVIDIA's new ASR model built on a fast conformer architecture, which achieves 2-3x inference speedups through depth-wise separable convolutional downsampling and CUDA graph optimizations. Looking ahead, Conway envisions a future of digital employees or agentic AI systems that augment human workforces, with Nemotron positioned for enterprise use cases requiring complex tool calling, on-premise deployment, data sovereignty, and multi-model orchestration. He highlights NVIDIA's European partnerships to bring reasoning capabilities to non-English languages and notes upcoming research directions including hybrid SSM/Mamba architectures and mixture-of-experts models for large-scale inference efficiency.

AI Radar wrote this summary from the publisher’s notes with AI, not from a transcript. How it works.