Physix Frontier · News Briefing Card (Leiphone · Nov 1, 2026)

Pika open-sources H3 Flex: 6-step joint audio-video generation

KEY FACTS

Pika released H3 Flex, a derivative model built on MiniMax H3, opening up adapter weights and inference code.
The model uses six-step denoising to generate picture and sound simultaneously, and supports text, first/last frame, and multimodal reference inputs.
The release includes two LoRA adapters, natively supports 480P and 768P, and can generate 5- to 15-second videos with synchronized stereo audio.
Pika offers 1080P, 2K, and 4K enhancement through its API, and splits audio into five separate tracks.
MiniMax H3 passed 24 million downloads in its first month as an open-source release, with more than 300 public derivative models.

KEY DATA

6 stepsDenoising steps

480P/768PNative output resolution

5-15 secondsVideo duration

Over 24 millionH3 first-month open-source downloads

PHYSIX OBSERVATION

Six-step sampling compresses joint audio-video generation to the level where drafts can be screened in real time, and open weights plus ComfyUI integration lower the barrier to creation. What is really worth watching is the API-side 4K enhancement and five-track separation, which connect generation and post-production into a single pipeline; open-source models are turning from demos into production tools.

Source: Leiphone report