Hao AI Lab integrates Pika's open-sourced MiniMax H3 Flex into FastVideo, achieving up to seven times faster inference.
Hao AI Lab announces that MiniMax H3 Flex, open-sourced by Pika last week, has been integrated into FastVideo. It is up to seven times faster than the official open-source inference code. On a single NVIDIA GB200 GPU, it generates a 5-second 480p video with audio in just 5.8 seconds. The lab is releasing weights and full training recipes, including a FastVideo tutorial, Hugging Face weights (https://huggingface.co/FastVideo/Pika-H3-Flex-NVFP4-QAD), and blog links.
Continue looking at this topic
OpportunitiesTikTok AI product promotion with cumulative ten million views: how to find benchmarks, break down viral hits, and judge content direction ↗Open-source projectHow to verify GLB files exported by anymove when converting an AI video into a character animation? ↗Hotly discussed viewsAI is tasked with creating educational videos based on implementation plans: how do we avoid scenarios where the visuals look great but the information is wrong? ↗News and AnalysisAI video can now reach 10 minutes: Should creators extend duration or rewrite scripts first? ↗Return to original text
A short introduction is retained here; the full text from the source website is not reposted. For specific wording and subsequent updates, please refer to the original article.
Open AIHOT original