NVIDIA Launches OpenReasoning-Nemotron Models Based on DeepSeek for Local AI Solutions on Gaming PCs

NVIDIA has launched a series of AI models capable of logical reasoning and problem-solving in mathematics, science, and programming. All models can be run locally—without the need for cloud services or expensive servers—using standard gaming GPUs.

The OpenReasoning-Nemotron line includes four models with parameter counts of 1.5, 7, 14, and 32 billion. These models were developed through distillation from a large-scale model, DeepSeek R1 0528 (with 671 billion parameters), utilizing the Qwen 2.5 architecture. NVIDIA employed the NeMo Skills pipeline, generating 5 million solutions and conducting training exclusively through supervised fine-tuning—without utilizing RLHF. This approach has yielded significant results: the largest model with 32 billion parameters scores 89.2 on AIME24 and 73.8 on HMMT, while the smaller 1.5 billion model scores 55.5 and 31.5 respectively.

All four models are available on the Hugging Face platform and can serve as a foundation for further research, including reinforcement learning and adaptation for practical applications. The GenSelect mode is also supported, allowing the generation of multiple responses to a single question and the selection of the best one. In this mode, the 32B model shows results comparable to OpenAI’s o3-high across various math and programming benchmarks.

The key feature of Nemotron is its accessibility. Even without cloud infrastructure, these models are suitable for local computations, scientific work, and the development of custom AI systems, provided you have a home PC with a modern mid-range gaming GPU.

Delegate some of your routine tasks with BotHub! There’s no need for VPN access, and you can use a Russian card. Follow this link to receive 100,000 free tokens for your initial tasks and start working with neural networks right away!