Pika releases H3 Flex — open video model with audio in 6 steps

Pika Labs has released the weights for Pika H3 Flex, a video generation model with audio fine-tuned on 33B MiniMax H3: generation takes 6 sampling steps instead of 50 — about 7 seconds for a 15-second clip in 768p.

Compiled by Sergey KostenchukPublished 2026-10-31Updated 2026-10-31

2026-10-31 Multimodal NVIDIA

🎨 Pika releases H3 Flex: open video model with audio in 6 steps

Pika Labs has fine-tuned 33B MiniMax H3 and released the weights for Pika H3 Flex — a video generation model with synchronized audio. Rendering has been reduced from 50 to 6 sampling steps: a 15-second clip in 768p is generated in about 7 seconds, a 1.8x speedup, with a price 80% lower than H3 Max.

🌍 Open weights and pricing starting from $0.01/sec lower the barrier to entry for studios and developers: drafts, up to 12 references (images, videos, audio), upscaling to 4K, and separate audio tracks (dialogue, music, SFX, foley, ambient) for post-production.

👤 The API is already live on dev.pika.art, and weights are available on Hugging Face for local deployment (requires an NVIDIA GPU, tested on H100 80GB). The model will appear in the pika.art app on November 2, 2026. According to the DesignArena preference aggregate, it ranks 1st among open-weights models.

Source 1: https://pika.art/blog/pika-h3-flex Source 2: https://github.com/Pika-Labs/Pika-H3-Flex

Sources