Pika H3 Flex
Pika's post-trained MiniMax H3 video model, distilled to 6 steps, with stereo audio and up to 12 reference inputs.
Overview
Pika H3 Flex, released October 29, 2026, is Pika's post-trained version of MiniMax's H3 video model, distilled from 50 to 6 denoising steps. It generates 5-15 second clips natively at 480p or 768p with synchronized stereo audio and accepts up to 12 reference images, videos or audio files. Generator adapters and inference code are open on Hugging Face and GitHub (Pika-Labs/Pika-H3-Flex); the upsampler to 1080p, 2K or 4K, five-stem audio separation (dialogue, score, SFX, foley, ambience), prompt enhancement and accelerated serving remain Pika-hosted. Pika says it ranked #1 among open-weights models and #7 overall on DesignArena's text-to-video leaderboard at launch.
This is a sourced directory entry. Features and capabilities are reported in the linked sources; this entry does not document independent Aymorphic testing.
Key features
- 6-step distilled sampling versus 50 steps for the base model (speed claim is transformer compute, not end-to-end latency)
- 480p or 768p native output, clips up to 15 seconds, synchronized stereo audio
- Up to 12 reference inputs (images, video, audio)
- Open weights for text-to-video, first/last-frame and reference-to-video adapters
- Pika-hosted upsampling to 1080p/2K/4K and five audio stems via the Pika API
- DesignArena text-to-video: #1 open-weights, #7 overall at launch (Pika-reported)
Use cases
- Fast iteration on short video drafts with reference-guided consistency
- Video with generated dialogue, score and sound effects
- Local or self-hosted video generation using the open adapters
Pricing
Unknown; not stated on the launch post. Available through the Pika API Club, and in the Pika creation product from November 2, 2026.
Pricing reflects the source record and may change. Confirm current terms with the provider.
Dates and sources
- Released
- 29 Oct 2026
- Record updated
- 1 Nov 2026
- Record last checked
- 1 Nov 2026