Pika H3 Flex: open-weights six-step video with sound, built on MiniMax H3, runs in ComfyUI
Pika H3 Flex is Pika's post-trained version of the open MiniMax H3 model, released with open weights on day one. It makes video with synchronized stereo audio in six sampling steps, takes up to 12 image, video and audio references, and runs locally in ComfyUI and diffusers.
What it does
It ships as two LoRA adapters on top of the MiniMax H3 base. FL2VA covers text to video and first-frame, last-frame or first-and-last-frame shots; Ref2VA takes up to 9 images, 3 videos and 3 audio clips (12 files in total) to carry identity, props, scene, motion and voice. Local output is native 480p or 768p, 5 to 15 seconds at 24 fps, in 11 aspect ratios, with 32 kHz stereo sound mixed in, and you can describe hard cuts, camera moves and dialogue for several shots in one prompt. ComfyUI has three ready templates (Text to Video, Reference to Video, First and Last Frame), and Comfy's guide says a 5-second 1344×768 clip finishes in about a minute on a single card. Pika says it debuted as the #1 open-weight model on DesignArena's text-to-video board.
Worth knowing
The adapters are under the MiniMax H3 Community License, so check its territorial and commercial-use conditions before using outputs commercially. In ComfyUI keep Pika's six-step schedule (lcm with ManualSigmas, CFG 1.0) and adapter strength 1.0, and do not stack a turbo or Lightning LoRA on top: it is already distilled, and more steps do not add detail. The Text to Video template needs ComfyUI 0.43.2 or newer. Comfy's testing notes say "wide shot" alone tends to render as a medium shot, so describe the subject as small in a large space, and timed camera moves can land 0.2 to 0.3 seconds off. Pika's own CLI was validated on an H100 80 GB and the full base components are about 210 GB; 1080p, 2K and 4K, separate audio stems and prompt enhancement are only on Pika's paid API.
Try it yourself
macro shot of a wooden tall ship in a violent storm, waves crashing over the deck, slow pull-back reveals the storm is inside a glass mug on a desk, warm lamp light, bookshelf bokeh