open-source

Pika opens the weights of H3 Flex, a video model with sound in six steps

Pika released H3 Flex on October 29, a post-trained version of MiniMax H3 that makes video with synchronized stereo audio in six sampling steps. Local weights do 480p and 768p; 1080p to 4K and separate audio stems stay on Pika's API.

Written by Mara at Sig · Oct 31, 2026

H3 Flex comes as two LoRA adapters on top of MiniMax-H3 base components: FL2VA for text plus first or last frame, and Ref2VA for up to 12 references (images, videos and audio clips). Pika cut generation from 50 denoising steps to six through distillation, and says Flex beats the models it tested on visual and audio quality. The press release says it ranks #7 for video quality on DesignArena's text-to-video board, above base H3. Those claims are Pika's.

The split is clear on the model card: local open-weight inference gives native 480p or 768p video with mixed stereo audio. Pika's hosted API adds 1080p, 2K and 4K, five editable audio stems (dialogue, score, SFX, foley and ambience) and prompt enhancement. The base components take about 210 GB on disk, and the card tests on a single H100 80GB. It runs in ComfyUI (BF16 LoRAs are included) and Diffusers. The license is the MiniMax H3 Community License, not Apache or MIT. In the app on pika.art, it arrives on November 2.

Why it matters

Six steps instead of 50 means less GPU time per draft clip, and the audio comes out with the video. You need a big GPU for local runs, and the resolution you'd ship to a client is API-only.

Try this

pip install huggingface_hub, then run hf download Pika-Labs/Pika-H3-Flex --include 'comfy/*' LICENSE NOTICE 'licenses/*' --local-dir ./pika-h3-flex-comfy and check the files with sha256sum -c SHA256SUMS inside the comfy folder, as Pika's ComfyUI guide shows.

Sources

Back to all news Get Sig Weekly every Monday