Pika has released the model weights for H3 Flex, reducing denoising steps from 50 to 6. This configuration generates 15 seconds of 768p video with stereo sound faster than real-time playback, utilizing 7 seconds of transformer computation time.
According to Pika, developers built a LoRA adapter based on the 33B MiniMax H3 model through distillation followed by reinforcement learning. Model distillation and open LoRA adapters reduce video generation costs, making local execution viable for developer infrastructure.
Running workloads locally requires significant computing resources. Processing these data volumes on proprietary hardware demands high-performance GPU clusters with high memory bandwidth, driving substantial capital expenditures for hardware and power. Precise inference cost figures for running H3 Flex on internal infrastructure remain absent from public sources and developer documentation.
Local deployment is restricted to 480p and 768p resolutions at lengths between 5 and 15 seconds. Resolutions from 1080p to 4K and five separate audio tracks remain exclusive to Pika's API, shifting heavy compute loads to the provider's infrastructure.
Commercial deployment of H3 Flex is governed by the MiniMax community license, which requires review prior to launching commercial operations.
The catch
4K rendering and five dedicated audio tracks remain locked behind the provider's API, while the company leaves the exact infrastructure costs for local GPU clusters out of its documentation. Will engineering teams invest in costly hardware for open 768p weights under MiniMax's community license while maximum 4K quality stays gated behind a proprietary API?
Our stories: MiniMax-H3: Native 2K Video and Audio in One AI Model (August 3).