The Efficient-Large-Model team—a joint effort between NVIDIA and MIT—has released the model weights and accompanying inference code for its video enhancement project, Open-H3-Regenerate-2K. Built on MiniMax H3, this solution regenerates low-resolution video drafts into high-resolution videos at up to native 4K. On the training side, the team added only a refinement LoRA, which, combined with community adapters, restores details through three-step denoising. On the inference side, optimizations such as latent-space upscalers and sparse attention have cut the time for a single NVIDIA B200 GPU to process a 5-second 2K video to 22.45 seconds—a nearly 3x speedup. Evaluations show that its 2K output scores higher in perceptual quality than the official corresponding API, falling slightly short only on metrics such as motion smoothness. Since MiniMax H3 opened its weights in August this year, its ecosystem has grown rapidly, surpassing 24 million downloads and spawning hundreds of public models. Many teams at home and abroad have already rolled out derivative work around it, including distilled versions, fine-tuned models, and adaptation tools.