Pika opened H3 Flex on October 29, 2026 — a six-step video-and-audio model post-trained from MiniMax H3, with reference-driven control (up to nine images, three videos, three audio clips) and 32 kHz stereo sound. It generates and conditions clips; it does not caption, reframe, or publish them.
2026-11-05 · by Moe Ameen
On October 29, 2026, Pika released H3 Flex, an open-weight video-and-audio model, on its API and on Hugging Face and GitHub; it reached the pika.art creation app on November 2, 2026. Rather than training from scratch, Pika post-trained the model from MiniMax H3 (the model MiniMax sells to consumers as Hailuo) and shipped it as LoRA adapters on roughly 33 billion parameters of base weights, with example inference code released under Apache 2.0. Pika's headline claim is speed: the model samples in six denoising steps instead of the teacher model's fifty, which Pika says yields 15 seconds of 768p video "in as little as 7 seconds" on an H100 — a controlled transformer-time benchmark, not an end-to-end API latency figure. Pika also claims it runs up to 1.8x faster and up to 80% cheaper than its H3 Max tier, averaged across text, image, and reference jobs.
The feature Pika is leading with is control. H3 Flex supports four modes — text-to-video, first-and-last-frame, reference-to-video, and reference-guided editing — and the reference mode is the headline: a single shot can be conditioned on up to twelve inputs, specifically nine images, three videos, and three audio clips, with at least one image or video required (audio-only requests are refused). Native output is 480p or 768p at 24 fps for clips of 5 to 15 seconds, across eleven aspect ratios from 21:9 to 9:21, and the model generates synchronized 32 kHz stereo audio with the video. Higher resolutions (1080p, 2K, 4K) and separated audio stems stay on Pika's hosted service.
The open weights cover inference for text-to-video, first-and-last-frame, and reference-to-video with synchronized audio at 480p and 768p. Several pieces deliberately stay hosted: the 4K upscaler, audio-stem separation, prompt enhancement, and Pika's accelerated serving stack. At announcement, Pika said H3 Flex ranked seventh on DesignArena's crowd-voted text-to-video leaderboard and led the open-weight entries there. Treat the performance and pricing figures as Pika's own, measured on its hardware; confirm current specs and plans on Pika's pages.
The practical consequence of a cheap, fast, open-weight generator is more clips, not fewer — and a pile of 768p MP4s is not a content week. That gap is where [Kompozy](/) does the work. Treat an H3 Flex render (self-hosted or via Pika's API) as source footage and drop it into Kompozy: run it through [Clipped Shorts](/glossary/clipped-short) to auto-caption and reframe one clip for every feed's aspect ratio, use it as the demo footage behind a hook in a [Marketing Short](/glossary/marketing-shorts), or set it as the portrait background under a Listicle Video's title-and-bullet cards. One render becomes several native edits instead of a single manual upload.
Then Kompozy does the multiplication the model can't. Under one [Persona Brief](/glossary/persona-brief) that locks your voice and banned words, the idea behind that clip also becomes a Text Post, brand-exact [Carousel Posts](/glossary/hyperframes), Quote Graphics, a Blog Article, and an Email Newsletter — and on the weeks you don't want to prompt a render at all, [Persona Shorts](/glossary/persona-shorts) generate net-new face-locked avatar video with no footage. [Autopilot](/glossary/autopilot) then schedules and publishes the whole set across the eight social platforms plus blog and email behind a per-post review gate. Generate the shot with H3 Flex; ship the content with Kompozy.
Pika H3 Flex is an open-weight AI video-and-audio model Pika released on 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, generates 5–15 second clips at 480p or 768p with synchronized 32 kHz stereo audio, and supports reference-to-video conditioning on up to 12 inputs. Higher resolutions and stem separation stay on Pika's hosted service.
The weights and example inference code are open — the adapters are on Hugging Face and GitHub, with the sample code under Apache 2.0 — so you can self-host text-to-video, first-and-last-frame, and reference-to-video at 480p and 768p. But the 4K upscaler, audio-stem separation, prompt enhancement, and Pika's fast serving stack remain paid hosted features, so "open" does not mean the full hosted pipeline is free. The weights themselves carry the MiniMax H3 Community License, not Apache 2.0, which excludes the US, EU, UK, and South Korea from local deployment without separate authorization — confirm the license file on the model page before self-hosting.
Up to twelve per shot in reference-to-video mode: nine images, three videos, and three audio clips, with at least one image or video required (audio-only prompts are refused). That multi-reference conditioning is the feature Pika is leading with, and it is what lets you hold a subject, style, or motion consistent across several generations.
No. H3 Flex is a generative model — it renders and conditions a clip and exports an MP4. It does not caption for feeds, reframe per platform, build carousels or blogs, or schedule and publish. To turn an H3 Flex clip into captioned posts across platforms, run it through a content engine like Kompozy, which captions, reframes, repurposes, and publishes across the eight social platforms plus blog and email.