An open-weight, audio-native AI video model post-trained from MiniMax H3 — six-step generation, synchronized 32 kHz stereo sound, and reference-driven control from up to nine images, three videos, and three audio clips per shot.
Last verified · 2026-11-05 · by Moe Ameen
Pika H3 Flex is an open-weight video-and-audio model Pika released on October 29, 2026, and brought to the pika.art app on November 2. Instead of training from scratch, Pika post-trained it from MiniMax H3 (sold to consumers as Hailuo) and published it as LoRA adapters on roughly 33 billion parameters of base weights, on Hugging Face and GitHub, with example inference code under Apache 2.0. The design goal is speed and control at a lower cost than Pika's hosted H3 Max tier.
The speed comes from sampling in six denoising steps rather than the teacher model's fifty. Pika reports 15 seconds of 768p video "in as little as 7 seconds" on an H100 — a controlled transformer-time benchmark, not an end-to-end latency promise — and claims up to 1.8x faster, up to 80% cheaper generation than H3 Max across text, image, and reference jobs. Native output is 480p or 768p at 24 fps, in clips of 5 to 15 seconds, across eleven aspect ratios from 21:9 to 9:21, with synchronized 32 kHz stereo audio generated alongside the video.
The headline feature is reference-driven control. H3 Flex has four modes — text-to-video, first-and-last-frame, reference-to-video, and reference-guided editing — and the reference mode conditions one shot on up to twelve inputs: nine images, three videos, and three audio clips, with at least one image or video required. That is what lets you steer subject, scene, motion, and sound rather than hoping a text prompt lands. Prompts run up to 8,000 characters and seeds are controllable.
H3 Flex is a model, not a studio. The open weights cover text-to-video, first-and-last-frame, and reference-to-video with audio at 480p and 768p; the 4K upscaler, audio-stem separation, and prompt enhancement stay on Pika's paid hosted service. It renders and conditions clips — it does not caption for feeds, reframe per platform, build carousels or blogs, or publish anywhere. Treat the performance and pricing figures as Pika's own, and confirm current specs on its pages.
H3 Flex's real advantage isn't the six-step speed — it's that you can condition a shot on your own references and hold a subject, style, and sound consistent across a whole batch. That makes it a brand-b-roll generator, not just a clip toy. But consistency inside the model stops at the MP4: H3 Flex won't wrap those clips in your brand frame, pair them with on-brand copy, or get them onto a feed. [Kompozy](/) is the layer that takes a consistent clip library and turns it into consistent, published posts.
The tightest fit is compositing. Feed an H3 Flex clip into [Persona Frames](/glossary/persona-frames), where Kompozy composites video as a movable layer inside a brand-exact [HyperFrames](/glossary/hyperframes) template — so your reference-locked footage lands inside the same frame, logo, and type as everything else you publish. The same clip can back a [Marketing Short](/glossary/marketing-shorts) hook or run through [Clipped Shorts](/glossary/clipped-short) for per-platform captioning and reframing. Then, under one [Persona Brief](/glossary/persona-brief) that governs voice and banned words, that visual idea extends into formats the model never touches — Carousel Posts, Quote Graphics, a Blog Article, and an Email Newsletter — and [Autopilot](/glossary/autopilot) schedules and publishes the whole set across the eight social platforms plus blog and email behind a per-post review gate. H3 Flex keeps the footage on-brand; Kompozy keeps the whole content operation on-brand and shipping.
Pika H3 Flex is an open-weight AI video-and-audio model released October 29, 2026, post-trained from MiniMax H3 and shipped as LoRA adapters on Hugging Face and GitHub. It samples in six steps for speed, makes 5–15 second clips at 480p or 768p with synchronized 32 kHz stereo audio, and can condition a shot on up to 12 references (9 images, 3 videos, 3 audio clips). Upscaling to 1080p–4K and stem separation stay on Pika's hosted service.
H3 Flex is Pika's open-weight, speed-and-control model built on MiniMax H3, where Pika's main app historically ran its own Pika model line with effects like Pikaffects and Pika Scenes. H3 Flex leads on fast six-step generation, synchronized audio, and multi-reference conditioning, and it can be self-hosted; the creative effects toolkit is a separate part of the pika.art product.
Yes, within limits. The weights are LoRA adapters on MiniMax H3 with open example inference code, so you can self-host text-to-video, first-and-last-frame, and reference-to-video with audio at 480p and 768p on your own GPU. The 4K upscaler, audio-stem separation, and prompt enhancement are hosted-only, so a fully local pipeline tops out at 768p. The weights themselves ship under the MiniMax H3 Community License, not Apache 2.0 — that license excludes the US, EU, UK, and South Korea from local deployment without separate authorization, so confirm the license file on the model page before self-hosting.
H3 Flex exports a clip; it doesn't publish. Run that clip through a content engine like Kompozy: composite it inside a brand-exact Persona Frames template, auto-caption and reframe it with Clipped Shorts, and fan the idea into a carousel, blog, and newsletter under one Persona Brief, then schedule and publish across the eight social platforms plus blog and email.