NVIDIA has introduced OpenReasoning-Nemotron, a new lineup of four distilled large language models purpose-built for reasoning tasks, scaled from 1.5 billion to 32 billion parameters. These compact models are derived from the massive 671-billion parameter DeepSeek R1 0528 model by applying knowledge distillation onto Qwen-2.5-based architectures. By leveraging the NeMo Skills pipeline, NVIDIA generated five million math, science, and coding solutions and used them to fine-tune each model via supervised learning—eschewing reinforcement learning or over-complicated methodologies. This approach results in models that excel in complex reasoning benchmarks, with the top 32B version notching an 89.2 on the AIME24 and 73.8 on the HMMT February contest, and even the smallest, the 1.5B variant, attaining respectable scores of 55.5 and 31.5 respectively.
The OpenReasoning-Nemotron models are designed as accessible research toolkits. All four versions will be openly downloadable from Hugging Face, providing a robust foundation for researchers to explore reinforcement learning-driven reasoning, or for developers wishing to tailor performance to specialized applications. Notably, the 32B model can be paired with GenSelect mode, which spawns multiple parallel generations for each question and selects the best answer; this pushes its performance to rival or even surpass OpenAI´s o3-high on a range of math and coding test suites. Given the purely supervised nature of fine-tuning, these models serve as clean, reproducible baselines for the community to conduct future reinforcement learning experiments without the complexities or potential biases of previous RL steps.
For the enthusiast and gaming segment, NVIDIA´s distillation strategy translates to top-performance reasoning capabilities that can run natively on higher-end consumer GPUs—sidestepping the need for cloud computing or costly enterprise-grade hardware. This move democratizes access to cutting-edge reasoning models, enabling individuals and smaller researchers to deploy, study, and modify state-of-the-art Artificial Intelligence solutions locally. OpenReasoning-Nemotron´s accessibility is poised to accelerate both applied and foundational Artificial Intelligence research, inviting widespread experimentation and innovation at every level of technical experience.
