Pika H3 Flex is Pika’s post-trained MiniMax H3 model for generating video with synchronized sound. Start with a prompt, guide the result with images and other references, and refine a short creative sequence. Released October 29, 2026; available through Pika API Club and downloadable adapters.
Flex produces picture and sound in six sampling steps. This makes it a candidate for exploring several directions before choosing a shot. Six steps describe the generation schedule; they do not promise a fixed completion time on every device or API request.
Supply an opening image, an ending image, or both, then describe the action between them. Endpoint conditioning suits a planned reveal or transition. Explain how the subject and camera move through the scene; the supplied frames are guidance, rather than a guarantee of exact motion.
Reference mode accepts images, video and audio together, with up to 12 files in total. Use a product photo for appearance and a clip for movement. The local package allows at most nine images, three videos and three audio clips within that shared limit, and requires at least one image or video.
Describe speech, physical sounds, ambience and music alongside the visual action. Local inference returns a mixed stereo soundtrack. For an editing workflow that needs separate layers, Pika’s hosted audio-stems service provides dialogue, score, sound effects, foley and ambience as separate tracks.
Generate the initial video at native 480p or 768p. After choosing a direction, Pika’s hosted upsampler can finish a draft at 1080p, 2K or 4K. High-resolution finishing is a separate hosted step; the downloadable adapters do not generate native 4K video.
Reference-guided prompts can describe a new character, setting or visual treatment while identifying elements to preserve. This is useful when developing variations from an established concept. Be specific about the change, then review identity, timing and continuity in the resulting clip; these controls express creative intent.
| Dimension | Pika H3 Flex · Hosted API | Pika H3 Flex · Local adapters |
|---|---|---|
| Access | Pika API Club | Downloadable adapters and inference package |
| Generation | Text, endpoint images and multimodal references, by task | FL2VA and Ref2VA task adapters on MiniMax H3 |
| Native video | 480p / 768p before the finishing pass | 480p / 768p |
| High-resolution finishing | Hosted upsampling to 1080p, 2K or 4K | Upsampler is not included in the local release |
| Sound | Synchronized audio; a separate service provides five stems | 32 kHz mixed stereo audio |
| Creative setup | Hosted serving and optional prompt enhancement | Your own hardware, dependencies and reference preparation |
| Choose it for | Trying creative directions through a hosted service | Evaluating or customizing an inference workflow |
These are two ways to use the same model family. Hosted finishing and audio-stem services are separate from the downloadable generation package.
Start with a product photograph and describe a reveal, a material detail or a camera move. Compare a few short treatments before committing to a longer edit. Inspect the product’s shape and details in every shot.
Use a character image alongside a movement reference to test a performance in a new setting. State which reference defines appearance and which defines motion. Review the generated action rather than assuming exact transfer.
Break a short scene into a few clear shots or describe one continuous take. Include the intended soundscape in the brief. Keep the action manageable within the requested duration, then check transitions and audible details.
Explore Pexo workflows for product films, cinematic scenes and animated stories. The previews below come from their Pexo pages.
Pika H3 Flex is Pika’s post-trained derivative of MiniMax H3, released on October 29, 2026. Its released adapters use a six-step sampling schedule for video with synchronized audio. It is a distinct Pika release, rather than another name for the original MiniMax H3 model or H3 Max.
The local package supports native 480p and 768p output at 24 fps, with 32 kHz stereo audio in an MP4. Requested durations are integer values from 5 to 15 seconds; actual length is rounded to the model’s frame grid. Eleven aspect ratios include widescreen, square and vertical canvases.
Yes. Ref2VA accepts up to 12 reference files in total, subject to limits of nine images, three videos and three standalone audio clips. Include at least one image or video. Individual video and audio clips must last 2–15 seconds; combined video duration and combined standalone-audio duration must each stay within 15 seconds.
As of October 30, 2026, Pika lists API Club access and downloadable adapters. Its launch announcement schedules availability in the Pika creation product for November 2. Native generation is 480p or 768p; 1080p, 2K and 4K use Pika’s separate hosted upsampler. Hosted audio-stem separation and prompt enhancement are also separate services.
The adapters are distributed under the MiniMax H3 Community License Agreement, which includes territory, use and commercial conditions. An Apache 2.0 license on example code or another component does not make the model weights unrestricted. Review the current model license and service terms for your intended deployment.
Pika H3 Flex is Pika’s post-trained derivative of MiniMax H3, released on October 29, 2026. Its released adapters use a six-step sampling schedule for video with synchronized audio. It is a distinct Pika release, rather than another name for the original MiniMax H3 model or H3 Max.
The local package supports native 480p and 768p output at 24 fps, with 32 kHz stereo audio in an MP4. Requested durations are integer values from 5 to 15 seconds; actual length is rounded to the model’s frame grid. Eleven aspect ratios include widescreen, square and vertical canvases.
Yes. Ref2VA accepts up to 12 reference files in total, subject to limits of nine images, three videos and three standalone audio clips. Include at least one image or video. Individual video and audio clips must last 2–15 seconds; combined video duration and combined standalone-audio duration must each stay within 15 seconds.
As of October 30, 2026, Pika lists API Club access and downloadable adapters. Its launch announcement schedules availability in the Pika creation product for November 2. Native generation is 480p or 768p; 1080p, 2K and 4K use Pika’s separate hosted upsampler. Hosted audio-stem separation and prompt enhancement are also separate services.
The adapters are distributed under the MiniMax H3 Community License Agreement, which includes territory, use and commercial conditions. An Apache 2.0 license on example code or another component does not make the model weights unrestricted. Review the current model license and service terms for your intended deployment.
