Pika H3 Flex: what creators get locally, and what stays in the cloud

A new release built around revisions

Pika describes H3 Flex as a post-trained MiniMax H3 model. Its launch announcement says it accepts up to 12 reference files across images, video and audio. Access starts through Pika API Club; the creation product is scheduled to receive Flex on November 2, 2026.

That gives a creative team a concrete question to test: can the model preserve a product, character or visual direction while the surrounding shot changes? A polished sample is encouraging, but repeated revisions are what expose whether a tool fits a production brief.

Public record: Pika launch announcement — October 29, 2026

Local weights and cloud finishing are different packages

The official repository describes local inference at native 480p or 768p with mixed stereo audio. Pika’s hosted API adds higher-resolution output at 1080p, 2K and 4K, plus audio stems and prompt enhancement. Downloading the model does not reproduce that complete hosted workflow.

The repository links model weights, inference code and setup guidance. Review the applicable model and code licenses before deployment; downloadable weights alone do not establish unrestricted usage rights.

For an editor, the distinction changes the handoff. A mixed soundtrack and separately adjustable tracks create different revision costs. For a developer, compare the complete pipeline you need, including finishing and infrastructure, before estimating savings from self-hosting.

Public record: Official Pika H3 Flex inference repository

Five audio tracks, and pricing that depends on the route

Pika’s technical report lists five editable stems: dialogue, score, sound effects, foley and ambience. Its published comparison tables give starting prices of $0.01 per output second for text/image-to-video and $0.03 for reference-to-video. These are starting rates rather than a quote for every configuration.

The report says post-training uses supervised fine-tuning, distribution matching distillation and reinforcement learning. It also reports vendor-run quality evaluations. Promptwire has not reproduced those experiments.

For a trial, record the chosen route and finishing settings alongside the bill. Count discarded generations as part of the cost of an accepted shot. The useful comparison is what it takes to deliver one usable clip.

Public record: Pika technical report and starting-price tables — October 29, 2026

Seven seconds does not mean a seven-second download

Pika’s speed footnote measures transformer computation for a 15-second, 768p clip in a controlled benchmark. The cited seven seconds excludes the complete end-to-end API journey. References, finishing, stem separation, hardware and service overhead can change completion time.

The launch announcement also clarifies that native generation is 480p or 768p; the 4K finish uses upsampling. Teams evaluating a delivery specification should distinguish those stages.

A practical first test is a short product scene with one requested revision. Measure submission-to-download time, inspect the product details frame by frame and listen to each available track. That exercise is our suggested evaluation, not a claim that Flex has already passed it.

Public record: Pika launch announcement — October 29, 2026