Yuesong, a researcher at Video Rebirth, open-sourced HyperFlow, a LoRA that distills MiniMax-H3's 49-step diffusion trajectory into 8 steps using data-free flow self-distillation, with the base model as its only teacher. It cuts generation time roughly 3x on H200 GPUs and reaches real-time speed on an 8-GPU B200 cluster when combined with other optimizations. Code, weights, and demo comparisons are public, under the MiniMax H3 Community License.
An arXiv paper introduces Video DeltaNet (VDN), a video-native hybrid attention architecture combining local Softmax attention with a linear-memory branch using Video Delta Attention. Instantiated on MiniMax H3, it denoises a 14.3-second 768p video in 6.70 seconds on eight B200 GPUs — a 14.5x speedup over the 50-step dense baseline — targeting livestream video generation.
An arXiv paper introduces Video DeltaNet (VDN), a video-native hybrid attention architecture combining local Softmax attention with a linear-memory branch using Video Delta Attention. Instantiated on MiniMax H3, it denoises a 14.3-second 768p video in 6.70 seconds on eight B200 GPUs — a 14.5x speedup over the 50-step dense baseline — targeting livestream video generation.
A new arXiv paper presents Sol-H3, a full-stack inference pipeline accelerating the 33-billion-parameter MiniMax-H3 open-source video diffusion model on both cloud (NVIDIA GB200) and edge (DGX Spark) hardware. It combines a two-stage cross-resolution generation scheduler with a Recursive Self-Improvement kernel-search loop, achieving up to 30x end-to-end speedup and 20% lower memory.
A new arXiv paper presents Sol-H3, a full-stack inference pipeline accelerating the 33-billion-parameter MiniMax-H3 open-source video diffusion model on both cloud (NVIDIA GB200) and edge (DGX Spark) hardware. It combines a two-stage cross-resolution generation scheduler with a Recursive Self-Improvement kernel-search loop, achieving up to 30x end-to-end speedup and 20% lower memory.
An X post from the MiniMax H3 team reports that SGL Project combined with VDN-H3 now runs MiniMax H3 beyond 2× real-time denoising on eight NVIDIA B200 GPUs. The setup generates 14.4 seconds of 768p video in 9.0 seconds end-to-end after warmup, with no measured quality regression. It demonstrates how open models gain speed through community-built inference optimization.
An X post from the MiniMax H3 team reports that SGL Project combined with VDN-H3 now runs MiniMax H3 beyond 2× real-time denoising on eight NVIDIA B200 GPUs. The setup generates 14.4 seconds of 768p video in 9.0 seconds end-to-end after warmup, with no measured quality regression. It demonstrates how open models gain speed through community-built inference optimization.
MiniMax released H3, a unified multimodal model that generates 5–15 second video clips at native 2K resolution with stereo audio. SGLang Diffusion provides day-0 serving support, enabling local inference on consumer/workstation GPUs like 2× RTX 5090 or 1× RTX Pro 6000.
MiniMax released H3, a unified multimodal model that generates 5–15 second video clips at native 2K resolution with stereo audio. SGLang Diffusion provides day-0 serving support, enabling local inference on consumer/workstation GPUs like 2× RTX 5090 or 1× RTX Pro 6000.
Researchers introduce Parallel Decoding Distillation (PDD), a simplified and scalable trajectory-based method for distilling diffusion and flow models into few-step generators. PDD avoids the mode collapse and optimization difficulties of variational score distillation and adversarial losses, enabling faster image and video generation.
Researchers introduce Parallel Decoding Distillation (PDD), a simplified and scalable trajectory-based method for distilling diffusion and flow models into few-step generators. PDD avoids the mode collapse and optimization difficulties of variational score distillation and adversarial losses, enabling faster image and video generation.