MiniMax Drops Open-Weight H3 AI Model Generating Video With Audio!
While Silicon Valley giants keep their video generators locked behind expensive subscriptions, a surprise release from MiniMax handed local weights with native stereo audio directly to developers.
The Chinese startup MiniMax released open weights for its 33-billion parameter multimodal model H3 on Hugging Face, turning heads across the AI landscape. While industry leaders like OpenAI and Runway hoard their code inside closed APIs like dragon treasure, this release lets engineers run a native video and sound generator on local hardware.
The core system generates 4 to 15 second clips at 24 frames per second with 32 kHz stereo audio in a single pass. Unlike traditional pipelines that slap sound effects on top as a clumsy afterthought, H3 produces lip-synced audio directly alongside the visual render, supporting character dialogue across 11 languages.
Under the hood, the open-weight release specifically provides the H3-Base module for 768p generation, while the high-resolution 2K upscaler and initial context parser remain locked on company servers. The download comes in two flavors: one focused on text-to-video and frame interpolation, and a multi-reference version capable of ingesting up to nine images and three audio tracks simultaneously.
According to benchmarking firm Artificial Analysis, the architecture currently holds first place in video editing tasks and second in text-to-video generation. The released community license allows free local execution, research, and custom fine-tuning, though commercial usage remains capped for companies earning over 20 million dollars annually.
Open source models continue to erode the digital moats of big tech monopolies one unexpected weights drop at a time. The tech giants might soon find out that holding back their best algorithms only accelerates their own irrelevance.
Source: Hugging Face
Comments
This is where the magic happens: AI reads your discussion and rewrites the article based on the most interesting comments. Each strong comment adds points to the meter below. Once the meter is full, the article updates live — no page reload needed.