Pika releases H3 Flex adapters and keeps high-resolution finishing in its hosted stack

Pika released H3 Flex adapters alongside a separate hosted video finishing workflow.

**Dek:** Pika’s MiniMax H3 derivative generates synchronized video and audio from prompts and references. The downloadable release and hosted API offer different finishing options and carry different obligations.

Pika introduced H3 Flex on October 29 and published generator adapters and inference code for its post-trained version of MiniMax H3. It supports text-to-video, first- and last-frame conditioning, and generation guided by multiple media references. The public release covers synchronized audio and video at 480p and 768p. The company made Flex available through its API Club and announced a November 2 rollout for the Pika creation product. That announcement supplies a rollout date, while the downloadable artifacts are independently visible in the public repository. [Pika launch](https://pika.art/blog/pika-h3-flex) · [Public model repository](https://huggingface.co/Pika-Labs/Pika-H3-Flex)

What creators can run themselves

The published workflow loads a pinned base and the adapter for the conditioning mode, prepares text and media inputs, and assembles the decoded video and sound into one output. An animation team can try first- and last-frame conditioning when a storyboard defines the start and end of a shot. Reference-based generation serves a different purpose, such as carrying a character or visual cue into a new scene. The documentation distinguishes adapters for these workflows, so a reproducible evaluation should record which mode and model revision created each clip. [Inference repository](https://github.com/Pika-Labs/Pika-H3-Flex) · [Technical deep dive](https://pika.art/blog/pika-h3-flex-research)

The license follows the component. Pika releases its inference and loader code under Apache 2.0. The model adapters and MiniMax base components use the MiniMax H3 Community License Agreement. The associated Qwen encoder assets have their own Apache 2.0 terms. A company considering redistribution or a commercial deployment should review the model agreement for the pieces it actually plans to ship and keep the notices with the applicable artifacts. [Component and license notices](https://huggingface.co/Pika-Labs/Pika-H3-Flex/blob/main/THIRD_PARTY_NOTICES.md)

Where the hosted service differs

Pika describes its 1080p, 2K and 4K upsampler, stem separation, optional prompt enhancement and accelerated serving as hosted features. A local test at the release’s native resolution therefore answers a different question from whether a commercial production can deliver a finished 4K shot with separately editable sound. The company’s timing example also measures seven seconds of transformer computation for a 15-second clip at 768p in a controlled benchmark. Upload time, queueing, reference preparation, decoding and finishing all affect how long a creator waits for a result. [Pika launch](https://pika.art/blog/pika-h3-flex)

A practical first comparison would reuse the same cleared reference images, first and last frames, and shot descriptions across several generations. An editor can then measure which outputs preserve the intended action, how often synchronized audio needs replacement, and how many revisions it takes to reach an acceptable clip. Those questions determine whether the local adapters help with iteration and whether the hosted finishing steps are worth using for delivery.