Pika announced H3 Flex on October 29, 2026 — not a brand-new video model from scratch, but a post-trained version of MiniMax H3 built for professional iteration. Most
AI video tools are built around one perfect first try. Pika's bet is different: creators iterate, compare versions, and tweak — so the model should be fast enough to make that loop painless.
What H3 Flex actually does
H3 Flex generates video with synchronized stereo audio. It can make a 15-second clip at 768p in as little as seven seconds of transformer computation — Pika is careful to note that is compute time, not total end-to-end API latency. It works with up to 12 reference inputs (images, video, audio), supports text-to-video, first- and last-frame generation, and multimodal reference-to-video. Draft at 480p or 768p, then upscale a favorite to 1080p1080p 2K, or 4K.
The 50-steps-to-6 trick
The headline engineering move is distillation: Pika compressed MiniMax H3's 50-step sampling process down to six steps, then refined the faster model with reward and preference training. That is why the speed claim is real but also narrow — fewer steps mean faster drafts, and Pika says quality held up: at launch, Flex ranked #1 among open-weight models and #7 overall on DesignArena'stext-to-video leaderboard Those are Pika's own reported benchmark results on its prompt set, so treat them as a launch claim, not an independent verdict.
Audio you can actually edit
Through the Pika API, Flex can export five separate audio stems: Dialogue, Score, SFX, Foley, and Ambience. That matters more than it sounds. A single mixed soundtrack is hard to fix; separate stems let an editor swap the music, mute a sound effect, or re-record dialogue without regenerating the whole clip. For creators, that is the difference between a demo and a usable asset.
Open weights, hosted finishing
Pika released Flex's generator adapters and inference code, so developers can run it on their own hardware under the model's license terms. Local generation covers 480p and 768p with synced audio; 4K upsampling, stem separation, and acceleration stay on Pika's hosted API. API pricing starts at $0.01 per output second for text/image-to-video and $0.03 for reference-to-video, per Pika's technical report. Flex is available through the Pika API Club now and arrives in the Pika creation product on November 2 — that rollout is upcoming, not already complete.
Why it matters
AI video is shifting from "look what it made" to "how fast can I fix it." Speed, references, and editable audio are workflow features, not demo features. If Flex's six-step quality holds up outside Pika's own tests, it lowers the cost of experimenting — and that is where the interesting videos come from.
H3 Flex is not magic, it is a practical trade: fewer steps, faster drafts, more control. For anyone making AI video regularly, that is arguably the more useful kind of launch.
Comments
Post a Comment