Nvidia

The new 49B model tops open benchmarks with a 128K context window, tool-use capabilities, and single‑GPU efficiency. It's a signal that NVIDIA aims to lead in agent‑focused LLMs that actually run in production.

NVIDIA has launched OpenReasoning-Nemotron, a suite of open-source reasoning-optimized language models ranging from 1.5B to 32B parameters. Built from DeepSeek R1, these models aim for state-of-the-art performance in math, science, and code reasoning. The models are fully available on Hugging Face, with recipes for local inference and reinforcement learning fine-tuning. This release strengthens the open LLM landscape with powerful reasoning-focused models accessible for both research and production use.