Pika Labs, best known for hosted generative-video tools, has released Pika H3 Flex, a post-trained adapter package for the 33-billion-parameter MiniMax-H3 model. The model card specifies six-step video and synchronized-audio generation on a single Nvidia H100 80GB GPU. The downloadable weights use the MiniMax H3 Community License.
Shorter generation times make rapid prompt, motion, and reference testing more practical. Pika reports that a 15-second clip can render in about seven seconds on an H100, although actual latency depends on resolution, conditioning inputs, memory strategy, and file encoding.
Two adapters, one 33B base
Pika’s research post defines two task-specific adapters that modify the same pinned MiniMax-H3 transformer.
Both adapters ship as FP32 LoRAs with rank 128 and scaling alpha 128. A LoRA changes selected model layers through compact matrices, allowing Pika to adapt the base model without distributing another full 33B checkpoint. The package also includes a BF16 Comfy-format build for lower-precision ComfyUI workflows.