NVIDIA’s New OpenReasoning-Nemotron Models Make Advanced AI Reasoning Open to Everyone

NVIDIA just made a big move in open-source AI. Last week, the company released OpenReasoning-Nemotron. It’s a new family of compact but powerful models designed for complex reasoning across math, code, and science. You can grab them right now on Hugging Face.

These models don’t just work well. They set new performance records in their size range. The 32B version hits a pass@1 score of 89.2 on AIME24 and 70.2 on LiveCodeBench, without any fine-tuning. If you enable GenSelect, NVIDIA’s own boost mode, that same model can reach 96.7 on HMMT Feb 2025, which is absurdly high for open models.

Interestingly, these models are distilled from DeepSeek’s 671B R1-0528 giant, a massive Chinese AI system considered on par with O3 and Gemini 2.5 Pro. NVIDIA trained its smaller models using a curated set of five million reasoning samples produced by DeepSeek. The result is high-level reasoning in models as small as 1.5B, all the way to 32B.

Yes, it’s the same DeepSeek that got flagged by a US House Committee as a “national security threat” back in April. But geopolitical tensions aside, NVIDIA’s release underscores how globally tangled AI innovation really is.

Developers don’t need frontier-scale infrastructure anymore. These models run with fewer resources, integrate easily with tools like NeMo and TensorRT-LLM, and are licensed for commercial use. For teams building reasoning-heavy applications, this drops the barrier to entry in a big way.