Pika Labs Opens 33B H3 Flex Audio-Video Model on Hugging Face
TL;DR
- Pika Labs released H3 Flex on Hugging Face, a 33B-parameter post-trained MiniMax-H3 model that generates video with synchronized audio in six sampling steps.
- Native output is 480p or 768p at 24 fps, 5 to 15 seconds long, across 11 aspect ratios, with 32 kHz stereo audio in an H.264 MP4 container.
- The ref2va mode accepts up to 12 inputs (9 images, 3 videos, 3 audio clips); 1080p, 2K, 4K upscaling and audio stems stay behind the Pika API.
Pika Labs has posted the weights for Pika H3 Flex, a 33-billion-parameter video model, on Hugging Face. The model is a post-trained variant of MiniMax-H3 and generates video with synchronized audio in six sampling steps.
The model card describes it as "a post-trained H3 model, designed to give professional creators the flexibility and speed to explore with precision." Native output is 480p or 768p at 24 fps, running 5 to 15 seconds, across eleven aspect ratios from 21:9 down to 9:21. Audio comes out at 32 kHz stereo, muxed into an MP4 container with H.264 video and AAC audio.
Reference-driven generation is the headline feature. Through the ref2va adapter, a single prompt can take up to 9 images, 3 videos, and 3 standalone audio clips, 12 inputs in total, with combined video and combined standalone audio each capped at 15 seconds. Other modes cover text-only, first-frame, last-frame, and first-and-last-frame conditioning.
Testing is on a single GPU, with the card specifying H100 80GB. "Large requests may require `--cpu-offload` and substantial host RAM," it notes, and base components run about 210 GB on disk. The resolutions anyone would actually ship to a client, 1080p, 2K, and 4K, stay behind the Pika API, along with audio-stem extraction and prompt enhancement.
License is the MiniMax H3 Community License Agreement.
Originally reported by huggingface.co
Read the original article →Original headline: Pika Opens 33B-Parameter H3-Flex Video Model With Six-Step Audio-Video Generation Across 11 Aspect Ratios