On October 29, 2026, Pika released Pika H3 Flex, a video generation model post-trained on MiniMax H3 and built for the way professional creators actually iterate: draft fast first, then refine toward the version they want. Pika cut sampling from 50 steps to 6 and says the faster model matches or beats its 50-step teacher on several measured quality dimensions.
Seven seconds is transformer compute time, not the whole job
Pika says Flex can generate a 15-second, 768p clip with synchronized audio in as little as seven seconds of transformer computation. An official footnote adds that this is compute time in a controlled benchmark, not end-to-end API latency, and that reference inputs, higher-resolution finishing, audio-stem separation, and hardware all affect actual completion time. At launch, Flex ranked first among open-weights models and seventh overall on DesignArena's text-to-video leaderboard — a launch snapshot, not necessarily a lasting position.
Up to 12 reference files, and five audio stems delivered separately
Flex accepts up to 12 reference files in one generation, mixing sketches, character sheets, product photos, moodboards, voice samples, and motion videos, and creators can say what each reference should contribute: one for the character, one for the world, one for the movement. Reference-guided editing can swap an outfit, a prop, or the whole setting while preserving the original performance and timing. On sound, Flex generates video with synchronized stereo audio, and through the Pika API it can also export five separate stems — Dialogue, Score, SFX, Foley, and Ambience — so each part of the mix can be worked on separately.
Open weights have boundaries; high-resolution finishing stays hosted
Pika also released Flex's open weights: the generator adapters and inference code are on Hugging Face and GitHub, covering text-to-video, first- and last-frame generation, and multimodal reference generation, with synchronized audio at 480p and 768p locally. The 1080p, 2K, and 4K upsampler, audio-stem separation, optional prompt enhancement, and accelerated serving stack remain Pika-hosted features, and local performance depends on hardware. Flex will also arrive in the Pika creation product on November 2, so creators do not have to rely on the API alone.
