MiniMax has announced FastH3 v1, an open-weight video generation release that the project says can generate a 15-second, 768p video in 13 seconds. In the first post of a six-part X thread, @haoailab described it as a collaboration involving FastVideo, nuvalab and FastGen, and reported up to 14x speedup on NVIDIA Blackwell GPUs.

Original post on X

Load the post to view it as published on X. X may receive connection data.

View the original post on X ↗

An open acceleration recipe is part of the release

The announcement says FastH3 v1 is fully open, allowing the community to run and improve its acceleration recipe. It also highlights open weights and recipes as a way to share not only performance gains, but how those gains are achieved.

That framing suggests the release is intended as a starting point for collaborative optimization rather than only a closed demonstration of faster inference. Developers and researchers can therefore view the announcement as an invitation to work on the acceleration approach itself, subject to the materials made available by the project.

The headline result has a specific hardware context

The central performance claim combines video length, resolution and generation time: a 15-second, 768p clip in 13 seconds. Separately, the post reports up to 14x speedup on NVIDIA Blackwell GPUs.

The hardware qualification matters when interpreting the result. The announcement ties its reported speedup to Blackwell hardware, while the phrase “up to” describes a maximum rather than a universal generation rate across every GPU or workload.

Post-training is the next focus

The post places FastH3 v1 within a wider shift toward open-weight video models and argues that this shift creates a need for post-training. It says post-training can improve speed and quality, while open weights and recipes make the process itself more accessible to the community.

It also cites community effort around Minimax H3 Max, based on MiniMax AI, as part of this direction. The broader message is that progress in video generation will involve not just producing faster clips, but sharing the methods that help improve speed and quality over time.