NVIDIA Introduces OpenReasoning-Nemotron Local AI Models for Logic, Math, and Programming Tasks

NVIDIA has announced a series of OpenReasoning-Nemotron language models designed to solve problems in the exact sciences, computer science and programming. The main difference is the ability to run locally without connecting to cloud services. The models are designed for use on home PCs with powerful video cards.

The line includes four options: with 1.5 billion, 7 billion, 14 billion and 32 billion parameters. All of them were created using the distillation technique based on the large-scale DeepSeek R1 model, which contains 671 billion parameters. Qwen‑2.5 was used as the architectural basis. Training was carried out on 5 million examples of problem solutions generated using the NVIDIA NeMo Skills tool.

A distinctive feature is the rejection of reinforcement learning (RLHF). The models underwent only supervised training, thanks to which NVIDIA positions them as a clean starting point for research and customization.

According to testing results, the largest model (32B) scored 89.2 points on the AIME24 math exam and 73.8 points in the February HMMT tournament. The younger version (1.5B) scored 55.5 and 31.5 points, respectively.

NVIDIA OpenReasoning-Nemotron is available for download on the Hugging Face platform. All models support GenSelect mode – running multiple responses in parallel and choosing the best one. According to NVIDIA, this allows the 32B model to reach or even surpass the OpenAI o3-high level on some math and programming tasks.

The company sees the new series as a universal tool for researchers, teachers, and enthusiasts. Local launch and open learning process make the models a convenient basis for further training and adaptation to applied tasks.