Pika open-weights a six-step video model with ears

What it does

Pika H3 Flex is a post-trained version of MiniMax-H3 that generates 5–15 second video with synchronized stereo audio in six sampling steps. It ships as two task adapters: FL2VA for text and first/last-frame conditioning, and Ref2VA for multimodal references — up to 9 images, 3 videos and 3 audio clips, 12 files total. Output is native 480p/768p at 24 fps across 11 aspect ratios, delivered as H.264/AAC MP4.

The interesting bit

The reference system is the real draw: you can steer a shot with ordered image, video and audio references rather than just prose, which is what makes multi-shot storytelling and motion transfer plausible from one model. Audio is generated jointly with the video, not bolted on afterwards.

Key highlights

• Six generator evaluations per clip — built for iteration speed, not marathon renders
• First-frame, last-frame, or both-endpoints conditioning modes
• Ref2VA accepts up to 12 reference files with per-type duration caps (2–15s each, 15s combined)
• Runs fully offline once the ~210 GB of base files and 5.5 GB of adapters are downloaded
• Pika’s hosted API adds 1080p–4K upscaling, audio stems and prompt rewriting — separate from the local package

Caveats

• Tested on a single H100 80GB; H200 and B300 are untested, and max-size reference combos can exceed GPU memory
• The showcase videos are creative demos, not validation renders — exact reproduction isn’t guaranteed
• 4K and audio stems are API-only, not part of the open weights

Verdict

Worth a look if you have serious GPU capacity and want controllable, audio-synced video generation you can run yourself. If you were hoping for 4K offline on a consumer card, the README is politely telling you no.