FULL STORY

MiniMax H3: Real-Time Generation, Open-Sourcing, and Backlash

After testers found MiniMax's H3 Max generates video faster than real time, MiniMax open-sourced H3 with FastH3 acceleration. While benchmarks confirmed the speed gains, the community erupted in debate over FastH3's quality gaps and release ethics.

2026-08-28 ~ 2026-08-29 · 4 episodes · 23 posts

Episode 1 · MiniMax H3 Max Generates AI Video Faster Than Real Time (2026-08-28, 4 posts)

Ethan Mollick found that MiniMax's H3 Max can generate quality AI video via a web interface faster than the time it takes to watch it, and developers have made it interactive, generating video in real time from chat prompts, marking a milestone for real-time AI video.

Episode 2 · MiniMax Open-Sources H3 with FastH3 Bringing Near-Real-Time Video Generation (2026-08-29, 10 posts)

On 08-29, MiniMax released its open-source video generation model H3, and together with Hao AI Lab, NVIDIA, FastVideo, Nuvalab and others launched the acceleration solution FastH3 v1, achieving near-real-time video generation; model weights, training code, and deployment guides are all open-sourced.

Confirmed

  • Performance: generates a 15-second 768p video in 13 seconds; up to 14x speedup on NVIDIA Blackwell GPUs
  • Hardware: a single B200 takes 47 seconds to generate a 15-second video, claimed at a 3x real-time factor; 4x B200 achieves near-real-time generation
  • Features: supports text-to-video with synchronized audio, fully open source with training code and deployment guides
  • Technical approach: built with FastVideo and Nuvalab using FastGen technology, as a fine-tuned accelerated version of MiniMax H3

Unconfirmed

  • Future plans mentioned officially (Omni ref support, NVFP4 quantization) remain on the roadmap with no release date
  • Support for consumer RTX GPUs is not clearly addressed in the posts

Why it matters

  • Near-real-time video generation at the single-GPU level significantly lowers the deployment barrier for video models, and open weights plus training code let researchers and developers reproduce and build on it locally
  • The multi-party collaboration (MiniMax, Hao AI Lab, NVIDIA) on inference acceleration shows video generation shifting from cloud queuing toward local, real-time interaction

Episode 3 · Tests Show MiniMax H3 Max Generates Video Dramatically Faster at Same Quality (2026-08-29, 2 posts)

Benchmark tests show fal's post-trained MiniMax H3 Max generated a 15-second 768p video in just 18 seconds—far faster than Seedance 2.5—while maintaining quality comparable to the original model.

Episode 4 · FastH3 Quality Dispute: Benchmarks Diverge and Release Ethics Under Fire (2026-08-29, 7 posts)

A quality controversy erupted in the community over the FastH3 fast-generation mode of MiniMax H3. isidentical implemented the original H3 based on official specs and compared it with FastH3, finding a clearly visible quality gap—the original is noticeably better, and some FastH3 outputs are close to random pixels, raising questions about why an official degraded version would ship unnoticed. isidentical further criticized the fastH3 project for riding the open-source hype with deceptive results, failing to honestly disclose the quality degradation, and creating false hope; the authors responded that they were compute-constrained but insisted the model is usable.

Confirmed

  • isidentical's tests showed a clear quality gap between the original H3 (implemented from official specs) and FastH3, with the original being better
  • A comparison by Reddit user rmrfallfiles argued FastH3 on 12GB of VRAM can match a 25-step ComfyUI workflow, but the FastH3 samples there were taken directly from the official blog, and the default workflow was raised from 20 to 25 steps with extra settings
  • FastH3's authors responded that they were compute-limited but insisted the model is usable

Unconfirmed

  • How large the real quality gap between FastH3 and the original H3 actually is: different reviewers reach opposite conclusions, and the Reddit comparison differs greatly in sample sources and workflow settings, making its conclusions hard to trust directly
  • Whether MiniMax officially knew about and later fixed the degradation; no relevant information in the material

Why it matters

  • Fal engineer jfischoff noted that parts of the community pursue "speed at all costs" model optimizations, claiming several-fold speedups over the original model while the output quality is completely distorted—essentially random pixels; he noted image/video models are extremely hard to optimize, and his team spends 2-3x more time on evaluation and verifying optimization correctness than on kernel engineering
  • isidentical stressed a quality-first release ethic: don't claim breakthroughs with immature checkpoints, and don't sacrifice quality with sparse attention just to hit 0.25-second ultra-fast generation, while criticizing the industry's "YOLO-style" slapdash culture
  • The incident highlights the need for rigorous evaluation and honest labeling of quality degradation amid the speed-optimization craze in open-source video models