Pika launches open-source H3 Flex video model with real-time generation and 4K output

What Pika is claiming

Pika describes H3 Flex as an omni video model family built by its own team on top of MiniMax H3. The company says it combined fine-tuning and reward-based optimization on an internally curated dataset selected for aesthetic quality. According to Pika, the model delivers state-of-the-art results across visual and audio quality metrics and human preference evaluations, and outperforms leading competitors by up to 91% while staying competitively priced.

In Pika's own human evaluation study, H3 Flex ranks #1 for human anatomy and for artistic and aesthetic styles. In a companion post, the company adds that at the time of writing the model ranks #1 among open-weights models and #7 overall on DesignArena's text-to-video leaderboard.

These rankings and comparisons come from Pika itself and the source does not detail the evaluation methods.

Speed, resolution and audio

Pika says H3 Flex generates faster than real time. It defines real-time as generation latency below the length of the output clip. The company states the model can produce a 15-second clip with synchronized audio at 768p in as little as seven seconds of transformer computation. To get there, Pika post-trained MiniMax H3 to generate in six steps instead of 50, then refined the faster model for quality and creative control.

Output resolutions run from 480p to 4K. H3 Flex also exports five separate, editable audio stems: dialogue, score, SFX, foley and ambience. These come alongside the generated soundtrack, which Pika calls a first for a video model.

The model accepts up to 12 reference files in a single generation, combining images, video and audio. Pika gives examples such as a sketch, a character sheet, a product photo, a moodboard, a voice sample or a video of the movement a user wants. Users can specify what each reference should contribute.

Beyond generation, Pika lists downstream editing tasks including style transfer, movement transfer, object replacement, character swap and targeted edits.

Pricing and how the comparison looks

Pika published two comparison tables. For text and image-to-video, H3 Flex starts at $0.01 per output second and is marked as real-time, 4K-capable, open-weights and offering audio stems. The table lists HeyGen Speed, Quality and Ultra at the same $0.01 starting cost, though without 4K output, open weights or audio stems, and with real-time not disclosed. MiniMax H3 is listed at $0.08, Fal H3 Max at $0.0325, Fal H3 Turbo at $0.01625, Seedance 2.5 at $0.103, FLUX 3 Video at $0.06, Gemini Omni 1.1 Flash at $0.101 and Wan 3.0 at $0.03.

For reference-to-video, H3 Flex starts at $0.03 per second and is marked as supporting 4K output, open weights, multiple image, audio and video references, and uploaded-audio lip sync or redub. Pika notes that Gemini Omni 1.1 Flash supports image and limited video references but not uploaded audio references, and that Seedance 2.5 and Wan 3.0 add a video reference charge.

Availability and what it means

H3 Flex is available through Pika's API, ComfyUI and Diffusers. Model weights are on Hugging Face and code is on GitHub. Pika also offers customized training services, with enterprise inquiries directed to [email protected].

For creators and developers, the combination of open weights, multi-reference input and separate audio stems points to a model aimed at iterative work rather than one-off generations. Pika says its low latency and open-source availability support quick iteration from concept to high-resolution output and let teams integrate the model into their own workflows. Its performance and price claims are the company's own and rest on its published tables and evaluations.