Pika introduced H3 Flex on October 29, 2026, positioning its post-trained MiniMax H3 model around creative iteration. The release combines prompts and media references with video and synchronized audio generation. API access and downloadable model artifacts are available; Pika says access in its creation product is coming November 2.
A reference is a direction, not a guarantee
Pika’s launch describes up to 12 reference files, mixing images, video and audio. A creator can supply a character, a setting and movement from different sources, then describe each source’s role. The more consequential idea is revision: preserve an existing performance while changing the surrounding scene or a prop.
For a marketing team, that could make it easier to explore campaign treatments around approved assets. Our assessment is that reference handling should be judged on what remains stable between revisions. If changing the background also changes a product’s proportions or a performer’s timing, a superficially attractive output may still create extra editing work.
Six generation steps and a narrower speed claim
The technical report describes supervised fine-tuning, distillation from 50 denoising steps to six, and subsequent reward and preference training. Pika also reports an optimized serving stack. These are different contributors to speed; downloading the adapters alone does not reproduce every hosted optimization.
The headline seven-second example measures transformer computation for a 15-second, 768p clip under controlled conditions. It is not total API turnaround. Media preparation, service overhead and finishing affect completion time. The company’s leaderboard and preference claims are launch snapshots and vendor-reported results, rather than evidence of universal superiority.
Production teams should measure the time from submitting a usable brief to approving an asset. That includes failed generations, downloading files, correcting sound and reviewing product fidelity. A faster first attempt is useful, but a faster approved sequence is the metric that changes a production schedule.
Downloadable generation and hosted finishing differ
The model card specifies native 480p and 768p output, 24 frames per second, and requested durations from 5 to 15 seconds. Reference mode accepts up to nine images, three videos and three standalone audio clips, subject to a 12-file total and additional duration limits. These are accepted input limits, not a promise that every combination fits available GPU memory.
Higher-resolution finishing, separate audio stems and optional prompt enhancement are hosted services. Local generation produces mixed stereo audio. The release’s generator adapters are MiniMax H3 derivatives under the MiniMax H3 Community License Agreement; the Apache license on example code does not apply to the model weights.
Where this could change a production workflow
Consider a short product sequence needing the same object across several camera angles. A useful workflow starts with approved product references and a simple shot plan. Generate alternatives at native resolution, review the object’s shape and movement, and spend finishing time on selected versions. Keep the source assets and revision decisions together so an editor can explain which reference informed the final clip.
This is an evaluation proposal, not a claim that Nerova has tested Flex. Teams should include difficult changes, such as replacing one prop while preserving hand contact or carrying a character through a cut. Those reveal whether reference control reduces work or merely shifts it into cleanup.
Pika’s own technical discussion leaves physics, consistency across scenes and complex reference control as areas for improvement. That makes H3 Flex most interesting as an iteration tool whose strengths can be measured against a real brief. The October release expands creative control and self-hosting options, while leaving finishing, rights review and final editorial judgment as explicit responsibilities.