NVIDIA and MIT Release Lightning OPD Framework, Boosting Model Distillation Efficiency 4x While Eliminating GPU Memory Issues

According to reports, NVIDIA and MIT researchers released Lightning OPD (Offline On-Policy Distillation), a new post-training framework for large language models that eliminates the need to keep a teacher model running during training. By precomputing the teacher model’s log-probabilities offline, the framework improves training efficiency by 4x while freeing all GPU resources for student model training.

In testing on 8 NVIDIA H100 GPUs, Lightning OPD successfully distilled Qwen3-30B-A3B-Base (a 30-billion parameter MoE model) and achieved 71.0 on the AIME 2024 benchmark, whereas standard OPD ran out of memory on the same hardware. For the smaller Qwen3-8B model, the framework required only 30 GPU hours to reach 69.9 points.