Reportedly: Pika open-sources H3 Flex, a MiniMax H3 derivative for joint audio-video generation in six steps
Model: MiniMax H3 · availability, license and releases
Reported by Leiphone and QbitAI · not yet confirmed by the company or a second independent outlet. We update this page when it is.
Pika released H3 Flex, a MiniMax H3 derivative with open adapter weights and inference code, Leiphone reported. It generates 480P/768P video with synchronized stereo audio in about 5–15 seconds using six-step denoising. Two LoRA adapters support text/first-last-frame and multimodal reference inputs.
China context
- Outside China
- Open weights · leiphone.com
- Claims
- Company-reported; not yet independently evaluated
- For builders
- Developers outside China can integrate H3 Flex into existing tools via ComfyUI and Hugging Face Diffusers, and run local inference for 480P/768P audio-video generation without paying for API calls.
- For investors
- Pika's open-weight release of a MiniMax H3 derivative may pressure closed video-generation API providers, as developers gain a free local option for prototyping and only pay for high-resolution upscaling.
Translated from Chinese. Quotes and facts link to the original sources.
Pika published H3 Flex, a derivative of MiniMax H3, and released adapter weights and inference code. The model jointly generates video and synchronized stereo audio in six denoising steps, supporting text, first/last frame, and multimodal reference inputs. It natively outputs 480P and 768P video of about 5–15 seconds. Pika provides ComfyUI workflows and Hugging Face Diffusers integration. Higher-resolution 1080P, 2K, and 4K enhancement and audio stem separation are available through Pika's API. Pika reported that H3 Flex achieved the highest scores on five visual and audio quality metrics among evaluated models and ranked eighth on the DesignArena text-to-video leaderboard, the highest-ranked open-source video model there. The release follows MiniMax H3's open-weight launch in August 2026, which has led to over 24 million downloads and more than 300 public derivative models.
Pika fine-tuned MiniMax H3 on an internal audio-video dataset, then used projection distribution matching distillation (PDMD) to compress the teacher model's 50-step denoising to 6 steps. Reward feedback learning and direct preference optimization trained LoRA adapters for aesthetics, clarity, and instruction following; adapters are merged for inference without extra steps. Pika also optimized hosted inference: on 8 NVIDIA B300 GPUs, generating a 5-second 768P audio-video clip took 2.129 seconds, down from 3.409 seconds, about 1.6× faster.
Pika's open-weight H3 Flex gives developers outside China a deployable MiniMax H3 derivative for joint audio-video generation, reducing reliance on Pika's paid API for high-resolution output. Competitors offering closed video-generation APIs now face a free, locally runnable alternative that can be integrated into existing creative tools via ComfyUI and Diffusers.
For developers, H3 Flex provides a low-cost way to prototype audio-video generation locally at 480P/768P, then upscale selected drafts via Pika's paid API. For Pika, the open release may drive adoption of its paid high-resolution and audio-stem separation services.
A verifiable next signal is whether Pika releases the H3 Flex model weights on Hugging Face under a permissive license, and whether independent benchmarks confirm the reported quality scores. Another signal is adoption by third-party tools or platforms beyond Pika's own ComfyUI workflow.