Pika H3 Flex - Faster-than-real-time video model with open weights
by•
Pika H3 Flex is a video generation model post-trained on MiniMax H3, with open weights. Per Pika, it runs 1.8x faster than H3Max, makes a 15 second clip in as little as 7 seconds, and costs up to 80% less. It handles text, image and reference-to-video plus editing, with up to 12 references, native stereo audio and exportable stems. Draft at 768p, upscale to 4K. Use it via the Pika API or Hugging Face.
Replies
Pika H3 Flex is Pika’s post-trained take on MiniMax H3, released with open weights and built to run faster than real time.
Pika says it generates 1.8x faster than H3Max, and a 15 second clip can take as little as 7 seconds, at up to 80% lower cost. Those are averages across matched comparisons, and they measure DiT time rather than end-to-end latency. Local inference also won’t match the hosted API’s speed.
On quality, it ranks #7 for video quality on the DesignArena text-to-video leaderboard (score 1293). Per Pika, that puts it above base H3, H3Max and Seedance 2.5.
What you get:
text-to-video, first and last frame image-to-video, and reference-to-video
up to 12 references (images, video, audio, sketches, moodboards)
prompt-based editing to add, remove or swap things in an existing video
draft at 768p, then upscale to 2K or 4K
native stereo audio, plus up to five exportable stems (Dialogue, Score, SFX, Foley, Ambience)
clips of 5 to 15 seconds at 24 fps, in 11 aspect ratios
Pricing is per output second. At launch, 768p text and image generation is $0.02 per second (regular $0.04) and 4K is $0.08 (regular $0.16). Reference-to-video costs more. There’s also a promo of 15 free generations a day for Pika API Club members, October 29 to November 5.
Good fit if you want a fast video model you can run through the API or download and run yourself, and you like drafting cheap before you upscale.
Try it: Pika H3 Flex · Pika API · Hugging Face weights