MetaPatchAI home
AI

Pika H3 Flex: Pika's New Video Model Is Fast, Takes 12 References, and Ships Open Weights

Compiled by MetaPatchAI · Published , 3:25 PM ET

Oct 30, 2026: Pika wants AI video to feel less like a slot machine and more like editing. On Oct 29, 2026, the company announced Pika H3 Flex, a video-and-audio model that Pika says can generate "faster than real time," take up to 12 reference files in one generation, export five separate audio stems, and finish clips at up to 4K. Pika also published open weights on Hugging Face the same day.

What Pika H3 Flex is

Flex isn't trained from scratch. Pika calls it "our post-trained MiniMax H3 model." The company says it:

  • Post-trained MiniMax H3 to generate in 6 steps instead of 50, then refined the faster model for quality and creative control.
  • Can generate a 15-second clip with synchronized audio at 768p in as little as seven seconds of transformer computation.
  • Ranked #1 among open-weights models and #7 overall on DesignArena's text-to-video leaderboard "at the time of writing."

Read these numbers with care. Pika's own footnote says the seven-second figure "measures transformer (DiT) computation for a 15-second 768p clip in a controlled benchmark, not total end-to-end API latency." The leaderboard spot "reflects the launch snapshot." These are Pika's own claims. We haven't run the model or checked the leaderboard ourselves.

The features creators will care about

Up to 12 references, mixed media. Flex accepts images, video, and audio together in one generation: a sketch, a character sheet, a product photo, a moodboard, a voice sample, or a motion clip. Pika says you can "tell it what each reference should contribute," for example a character from one image, the world from another, and motion from a video.

Reference-guided editing. If part of a shot works, you can replace an outfit, a prop, or the background while asking Flex to keep the performance and timing.

Audio stems. Flex makes video with synchronized stereo audio. Through the Pika API, you can also request five stems: Dialogue, Score, SFX, Foley, and Ambience.

Draft low, finish high. Flex generates natively at 480p or 768p. Pika's upsampler then takes a chosen clip to 1080p, 2K, or 4K.

What "open weights" means here

This is the part to read closely. Pika says it's releasing "Flex's generator adapters and inference code." The Hugging Face model page (Pika-Labs/Pika-H3-Flex, created Oct 29, 2026) lists:

  • Base model: MiniMaxAI/MiniMax-H3, with the relationship marked as "adapter" and a "lora" tag
  • License: "other," named minimax-h3-community-license-agreement
  • Library: diffusers

The open release covers text-to-video, first- and last-frame generation, and multimodal reference-to-video, with synchronized audio at 480p and 768p. Pika says these stay Pika-hosted only:

  • the 1080p / 2K / 4K upsampler
  • audio-stem separation
  • optional prompt enhancement
  • the accelerated serving stack

So the local version won't have 4K finishing or stems, and Pika notes that "local performance depends on your hardware and setup."

Where to try it

  • Pika API (through what Pika calls the Pika API Club), available now per the announcement.
  • Pika's creation product: Flex is "coming to the Pika creation product on November 2."
  • Self-hosted: weights on Hugging Face, code on GitHub (Pika-Labs/Pika-H3-Flex).

What's confirmed / not confirmed

Confirmed (Pika blog, Oct 29, 2026; Hugging Face model metadata): - Pika H3 Flex is a post-trained MiniMax H3 model, released Oct 29, 2026. - It takes up to 12 references and outputs synchronized stereo audio. Five stems are available via the API. Native output is 480p/768p, with hosted upscaling to 4K. - The open release is adapters plus inference code on top of MiniMax H3, under a MiniMax H3 community license. - It comes to the Pika creation product on Nov 2.

Not confirmed / Pika's own claims: - The "faster than real time" and seven-second figures are Pika's benchmark (DiT compute only), not independently tested. - The DesignArena ranking (#1 open-weights, #7 overall) is Pika's launch snapshot. We didn't check the live leaderboard. - Pricing for Flex through the API wasn't in the announcement. - What the MiniMax H3 community license allows for commercial use. Read the LICENSE file before you ship anything.

Sources

Related

Up: AI