Pika Labs has released Pika H3 Flex, two open-weight LoRA adapters that reduce MiniMax H3’s 50-step sampling schedule to six steps. The adapters preserve H3’s single-pass generation of synchronized video and audio. Pika reports that a 15-second clip requires roughly seven seconds of transformer compute on an H100 80GB GPU.

Developers already running MiniMax H3, the 33-billion-parameter video-audio model also called Hailuo 3, can reuse the existing base checkpoint and architecture. Flex adds a task-specific weight delta and an exact sampling schedule, avoiding another full 33B model download.

A small delta on a 33B base

LoRA adapters store compact, low-rank updates to a larger model’s weights. Pika trained its adapters on a curated mix of video, audio, and reference-conditioning tasks, distilled the generator from 50 steps to six, and refined the result with reward feedback and preference optimization.

Both packages use FP32 weights with rank 128 and alpha 128, which define the size and scaling of the low-rank update. Pika also publishes BF16-converted weights for ComfyUI users.

Local output stops at 768p

Local inference supports native 480p and 768p output across 11 aspect ratios ranging from 21:9 to 9:21. Clips run at 24 fps, last between 5 and 15 seconds, and include 32 kHz stereo audio mixed into the MP4.