Pika Labs open-sources H3 Flex, a MiniMax H3 finetune with 6-step generation and 12 reference inputs
LatentSpacer · reddit · 2026-10-30
Pika Labs released Pika H3 Flex on Hugging Face, a post-trained finetune of MiniMax H3 for video generation.
- Built for professional iteration: synchronized-audio video in six sampling steps, up to 12 reference inputs for shot direction via references and prompts
- Local open-weight inference natively outputs 480p/768p with mixed stereo audio; Pika's hosted API extends to 1080p, 2K and 4K with audio stems and prompt enhancement
- Weights live at Pika-Labs/Pika-H3-Flex, with quickstart, prompt guide and GitHub repo
More from Multimodal
- SpAItial launches Echo-3: minutes of phone capture become real-scale 3D Gaussian Splat worlds — rmbrualla · 2026-10-30
- Native 3K MiniMax H3 video on a 4090: one ComfyUI node kills the VRAM crash — friedlc · 2026-10-30
- Pika H3 Flex launches: faster-than-realtime video model, 1.8x quicker and 80% cheaper than H3Max — minchoi · 2026-10-30
- 'Draft-then-upscale' will likely become the default image workflow soon — socialwithaayan · 2026-10-30
- Inworld launches Reasoning for Realtime TTS-2, first context-steering TTS model — Scobleizer · 2026-10-30
- Runway's App Studio turns short films into full video games, no coding required — tlakomy · 2026-10-30