Pika announced H3 Flex on 29 October. The post-trained MiniMax H3 model accepts up to 12 image, video and audio references in one generation, allowing a creator to guide character, setting, movement, composition and sound with source material rather than text alone.

Pika says H3 Flex can produce a 15-second 768p clip with synchronized audio in as little as seven seconds of transformer computation. That figure is a controlled vendor benchmark, not total API latency. The company reduced the generation process from 50 denoising steps to six; reference processing, upscaling, audio separation, hardware and service overhead can add time.

What it means for creative production

The most useful production change is the combination of references and separated sound. Through Pika’s hosted API, a result can include Dialogue, Score, SFX, Foley and Ambience as five individual stems. An editor can therefore change the balance without lowering every sound or regenerating the picture.

The open release covers text-to-video, first- and last-frame generation, multimodal reference-to-video and synchronized audio at native 480p or 768p. Pika keeps its 1080p-to-4K upsampler, stem separation, optional prompt enhancement and accelerated serving stack as hosted features. Local performance depends on hardware and setup.

A useful trial should use one short sequence with an approved character or product reference, a movement reference and a clear sound direction. Review continuity, reference fidelity, audio synchronization and how cleanly each stem can be remixed. Record total turnaround time rather than the model’s transformer time alone, and check the open-model licence and hardware requirements before planning local use. Unreal Studio has not independently benchmarked the release.

Primary sources

Pika · Announced 29 October 2026 · Verified 31 October 2026.

← Back to Intelligence