Pika Open-Sources H3 Flex Based on MiniMax H3, Achieving Joint Audio-Video Generation in Six Steps

Recently, Pika released Pika H3 Flex, a derivative model of MiniMax H3, and opened up the adapter weights and inference code. The model jointly generates visuals and sound through six-step denoising, supports text, first and last frame, and multimodal reference inputs, and is intended for creators to use in shot design, reference generation, and video editing.

This release includes two sets of LoRA adapters, one for text and first/last frame conditional generation and one for multimodal reference generation, along with the MiniMax H3 base components and local inference tools. The model natively supports 480P and 768P output and can generate videos of about 5 to 15 seconds with synchronized stereo audio. Pika provides a ComfyUI workflow and a Hugging Face Diffusers integration to make it easy for developers to add the model to their existing creative tools.

Beyond local generation, Pika can provide 1080P, 2K, and 4K video enhancement as well as audio stem separation via API. Creators can first generate and screen drafts at lower resolution, then process the selected version at high resolution; the synchronously generated sound can be split into five independent tracks—dialogue, music, sound effects, Foley, and ambience—for separate adjustment in post-production. The high-resolution enhanced video and audio stem separation are provided by the Pika API service.

In a technical blog post, Pika explained that the team first performed supervised fine-tuning of MiniMax H3 on an internal audio-video dataset, then used projected distribution matching distillation (PDMD) to compress the teacher model's 50-step denoising to 6 steps. The team then trained LoRA adapters through reward feedback learning and direct preference optimization to improve aesthetic quality, clarity, and instruction following. After the adapters are merged for inference, no additional generation steps are required.

Pika also accelerated its hosted inference service. In tests that kept the model weights and the six-step sampling scheme unchanged, generating a 5-second, 768P audio-video clip using 8 NVIDIA B300 GPUs reduced the combined time for condition encoding, denoising, and decoding from 3.409 seconds to 2.129 seconds, roughly 1.6 times the original processing speed.

In terms of visual and audio performance, the comparative evaluation results published by Pika show that H3 Flex achieved the highest scores among the evaluated models on all five visual and audio quality metrics, and ranked eighth on the DesignArena text-to-video leaderboard, making it the highest-ranked open-source video model on that leaderboard.

In August this year, after MiniMax H3 was open-sourced, MiniMax and partner companies continued to train and optimize the model around different use cases and released it to the community, including MiniMax H3 480P built with Running Hub; H3 Max 768P and H3 Max 480P with fal, which chose to do post-training and inference optimization for real-time generation; and FastH3 from FastVideo through sparsification and hardware adaptation, StreamRender-H3 and Open-H3-Regenerate-2K developed by NVIDIA and MIT jointly around H3, and several other models. Overseas developers have built innovative uses around H3 derivative models, such as Twitch live streaming, a 24-hour "AI TV station," and AI short-video live streaming software.

Within one month of being open-sourced, H3 has been downloaded more than 24 million times and has produced more than 300 public derivative models, making it one of the most downloaded models in the world in 2026.