Pika announces the launch of its brand-new open-source model, Pika H3Flex, powered by MiniMax H3. It enables high-quality audio-visual generation through six-step denoising and supports multiple input types, significantly enhancing shot design and audio-visual synchronization efficiency. Ranked eighth on the DesignArena text-to-video leaderboard, it is the highest-ranked open-source video model. Since the open-sourcing of MiniMax H3, it has been downloaded over 24 million times globally and has inspired more than 300 derivative models, placing it among the most active ecosystems worldwide.Author and source: AIBase
In recent years, the pace of iteration of artificial intelligence in audiovisual creation has been astonishing. Recently, the renowned AI video creation platform Pika announced the launch of a new derivative model based on MiniMax H3—Pika H3Flex—and has open-sourced the adapter weights and complete inference code, providing creators worldwide with a powerful new tool for shot design and multimedia content generation.
Six-Step Denoising with Multi-Class Inputs: Creating an Efficient Audio-Visual Joint Generation Experience
As a core technological highlight, Pika H3Flex innovatively enables high-quality audiovisual generation through just six denoising steps. In practical creation, the model supports multiple input methods, empowering creators to easily perform complex tasks such as flexible shot design, camera movement, and synchronized audiovisual effects.
In authoritative benchmark evaluations, the model has demonstrated exceptional technical strength. Comparative benchmark data shows that Pika H3Flex achieved the highest scores among all evaluated models on multiple core audio-visual quality metrics. On the renowned DesignArena text-to-video leaderboard, it rose to eighth place and became the highest-ranked open-source video model on the list.
Thriving Ecosystem: MiniMax H3 Ignites Global Open-Source Movement
Looking back at the underlying technology ecosystem, since MiniMax H3 was officially open-sourced in August this year, it has sparked widespread enthusiasm within the global developer community. Industry partners and overseas developers have conducted in-depth optimizations for various vertical application scenarios, successively launching multiple derivative models and giving rise to imaginative innovations such as Twitch streaming and AI television channels.
Official statistics show that within just one month after the open-source release of MiniMax H3, its cumulative downloads exceeded 24 million, and the number of publicly available derivative models surpassed 300, making it one of the most downloaded and actively utilized AI models globally in 2026. The addition of Pika H3Flex undoubtedly brings further strong innovative momentum to this open-source ecosystem.


