{"id":26232,"date":"2025-07-20T05:39:59","date_gmt":"2025-07-20T05:39:59","guid":{"rendered":"https:\/\/youzum.net\/nvidia-ai-releases-openreasoning-nemotron-a-suite-of-reasoning-enhanced-llms-distilled-from-deepseek-r1-0528\/"},"modified":"2025-07-20T05:39:59","modified_gmt":"2025-07-20T05:39:59","slug":"nvidia-ai-releases-openreasoning-nemotron-a-suite-of-reasoning-enhanced-llms-distilled-from-deepseek-r1-0528","status":"publish","type":"post","link":"https:\/\/youzum.net\/zh\/nvidia-ai-releases-openreasoning-nemotron-a-suite-of-reasoning-enhanced-llms-distilled-from-deepseek-r1-0528\/","title":{"rendered":"NVIDIA AI Releases OpenReasoning-Nemotron: A Suite of Reasoning-Enhanced LLMs Distilled from DeepSeek R1 0528"},"content":{"rendered":"

NVIDIA AI has introduced OpenReasoning-Nemotron<\/strong>, a family of large language models (LLMs) designed to excel in complex reasoning tasks across mathematics, science, and code. This model suite\u2014comprising 1.5B, 7B, 14B, and 32B parameter versions<\/strong>\u2014has been distilled from the 671B DeepSeek R1 0528 model<\/strong>, capturing its high-level reasoning capabilities in significantly smaller and more efficient models.<\/p>\n

The release positions NVIDIA as a leading contributor to the open-source LLM ecosystem, delivering models that push state-of-the-art (SOTA) performance while remaining commercially permissive and widely accessible via Hugging Face<\/a>.<\/p>\n

Model Overview and Architecture<\/strong><\/h3>\n

\"\u2705\" Distillation from DeepSeek R1 0528 (671B)<\/strong><\/h4>\n

At the heart of OpenReasoning-Nemotron lies a distillation strategy<\/strong> that transfers reasoning ability from DeepSeek R1\u2014a massive 671B parameter model\u2014into smaller architectures. The process prioritizes reasoning generalization<\/strong> over raw token prediction, enabling compact models to perform effectively on structured, high-cognition tasks.<\/p>\n

The distillation dataset emphasizes mathematics, science, and programming languages<\/strong>, aligning model capabilities with key reasoning domains.<\/p>\n

\"\ud83d\udcca\" Model Variants and Specs<\/strong><\/h4>\n
\n\n\n\n\n\n\n\n\n
Model Name<\/th>\nParameters<\/th>\nIntended Use<\/th>\nHugging Face Page<\/th>\n<\/tr>\n<\/thead>\n
OpenReasoning-Nemotron-1.5B<\/td>\n1.5B<\/td>\nEntry-level reasoning and inference<\/td>\nLink<\/a><\/td>\n<\/tr>\n
OpenReasoning-Nemotron-7B<\/td>\n7B<\/td>\nMid-scale reasoning, good for code\/math<\/td>\nLink<\/a><\/td>\n<\/tr>\n
OpenReasoning-Nemotron-14B<\/td>\n14B<\/td>\nAdvanced reasoning capabilities<\/td>\nLink<\/a><\/td>\n<\/tr>\n
OpenReasoning-Nemotron-32B<\/td>\n32B<\/td>\nNear frontier-model performance in logic-intensive tasks<\/td>\nLink<\/a><\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/figure>\n

All models are compatible with transformer architectures<\/strong>, support FP16\/INT8 quantization<\/strong>, and are optimized for NVIDIA GPUs and NeMo<\/strong> frameworks.<\/p>\n

Performance Benchmarks<\/strong><\/h3>\n

These models set new state-of-the-art pass@1 scores for their size class<\/strong> across multiple reasoning benchmarks:<\/p>\n

\n\n\n\n\n\n\n\n
Model<\/td>\nGPQA<\/td>\nMMLU\u2011PRO<\/td>\nHLE<\/td>\nLiveCodeBench<\/td>\nSciCode<\/td>\nAIME24<\/td>\nAIME25<\/td>\nHMMT\u202fFeb\u202f2025<\/td>\n<\/tr>\n
1.5B<\/td>\n31.6<\/td>\n47.5<\/td>\n5.5<\/td>\n28.6<\/td>\n2.2<\/td>\n55.5<\/td>\n45.6<\/td>\n31.5<\/td>\n<\/tr>\n
7B<\/td>\n61.1<\/td>\n71.9<\/td>\n8.3<\/td>\n63.3<\/td>\n16.2<\/td>\n84.7<\/td>\n78.2<\/td>\n63.5<\/td>\n<\/tr>\n
14B<\/td>\n71.6<\/td>\n77.5<\/td>\n10.1<\/td>\n67.8<\/td>\n23.5<\/td>\n87.8<\/td>\n82.0<\/td>\n71.2<\/td>\n<\/tr>\n
32B<\/td>\n73.1<\/td>\n80.0<\/td>\n11.9<\/td>\n70.2<\/td>\n28.5<\/td>\n89.2<\/td>\n84.0<\/td>\n73.8<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/figure>\n

All quoted scores are pass@1 without GenSelect.<\/p>\n

\"\ud83d\udd0d\" GenSelect (Heavy Mode)<\/strong><\/h4>\n

Using Generative Selection with 64 candidates<\/strong> (\u201cGenSelect\u201d), performance further improves, especially at 32B:<\/p>\n