Pika H3 Flex: Open-Weights AI Video in 6 Steps Instead of 50, With Synced Audio

Most AI video models make you wait through about 50 denoising steps for a few silent seconds. Pika's new Pika H3 Flex, announced on 29 October 2026, does the job in 6 steps instead of 50 — and it generates the soundtrack in the same pass. It is a post-trained version of MiniMax H3, released with open weights, so creators can run it themselves as well as use it through Pika.

What Pika H3 Flex actually is

Flex is not a brand-new model built from scratch. Pika took MiniMax H3 and post-trained it in three stages, set out in its technical deep dive: supervised fine-tuning on an in-house, curated audio-video dataset, then Projected Distribution Matching Distillation (PDMD) to cut sampling from 50 steps to 6, then reward learning — ReFL for aesthetics and clarity, and DPO for instruction following and realism. The open release is two generator adapters: FL2VA for text-to-video and first/last-frame generation, and Ref2VA for multimodal reference-to-video.

The verified specs

• Steps: 6 denoising steps, down from the teacher's 50 — roughly 8x fewer passes.
• Clip length: 5–15 seconds in the open-weights release, at 24 fps; a 15-second request produces 362 frames (about 15.08 seconds).
• Audio: synchronized stereo audio at 32 kHz, mixed into the video — dialogue, music and effects generated with the picture, in MP4 (H.264 video, AAC audio).
• Resolution: native 480p or 768p locally, across 11 aspect ratios from 21:9 to 9:21. On Pika's hosted service the same take can then be upscaled to 1080p, 2K or 4K.
• References: up to 12 reference files in one generation — up to 9 images, 3 videos and 3 audio clips — for characters, products, movement or voice.
• Ranking: at launch Pika said Flex ranked #1 among open-weights models and #7 overall on DesignArena's text-to-video leaderboard — that is Pika's launch snapshot, not an independent test by us.

How it compares with normal diffusion video

A normal diffusion video model refines noise into frames over dozens of sequential steps, so every extra second of video costs more time and compute, and the audio is usually added later in a separate tool. Flex was distilled to follow a fixed 6-step schedule, so it needs far fewer transformer passes per clip. Pika says a 15-second 768p clip with audio took as little as 7 seconds of transformer computation in its controlled benchmark. That is the headline — read the caveat below before quoting it as a real-world wait time.

How to try it, step by step

1. Easiest route — Pika itself: go to Pika's official H3 Flex announcement and follow its "Try Pika H3 Flex through the Pika API Club" link. Pika also said Flex was coming to the Pika creation product on 2 November 2026. Through the API you also get the hosted extras: upscaling to 4K and separate audio stems.
2. Download the open weights: the weights are on Hugging Face at huggingface.co/Pika-Labs/Pika-H3-Flex — this is the exact page linked from Pika's official blog.
3. Get the code: the inference code and setup instructions are on GitHub at github.com/Pika-Labs/Pika-H3-Flex, also linked from Pika's blog. Clone that repository and follow its Install section — it uses Python 3.12 and an NVIDIA GPU.
4. Start small: the GitHub guide itself advises starting qualification at 480p / 5 seconds, for example with its t2va text-to-video task and a prompt file, before trying 768p or 15 seconds.
5. Draft, then finish: generate cheap 480p/768p drafts, keep the take you love, and only then pay for the hosted 1080p, 2K or 4K finish if you need it.

The honest caveats

• The "7 seconds" is not your wait time. Pika's own footnote says it measures transformer (DiT) computation only, in a controlled benchmark. Reference inputs, 4K finishing, audio-stem separation, hardware and service overhead all add time on top.
• Open weights does not mean light. Pika's GitHub guide says downloading both tasks takes about 210 GB of base files plus 5.5 GB of adapters, and its representative tests run on one H100 80GB GPU per request. This is not a laptop model for most people.
• Some features stay hosted. The 1080p/2K/4K upsampler, audio-stem separation, prompt enhancement and the accelerated serving stack are Pika-hosted only. Locally you get native 480p/768p with mixed stereo audio. Note the stem count also differs by page: Pika's blog describes five stems via the API (Dialogue, Score, SFX, Foley, Ambience), while its product page lists three exportable stems (Dialogue, Music, Effects) — check the live page before you rely on either number.
• Licence terms apply. The adapters are model derivatives of MiniMax H3 and are distributed under the MiniMax H3 Community License Agreement — read it on the Hugging Face page before commercial use.
• Speed claims are Pika's. In its technical post, Pika's inference optimisations cut a 5-second 768p clip from 3.409 to 2.129 seconds of full inference on eight B300 GPUs. Impressive, but it is a datacentre benchmark, not a promise for a home GPU.

Why it matters for creators

AI video has been a waiting game: change one word in a prompt, then wait minutes to see if it helped. A model built for fast drafts with sound built in changes the workflow to real creative iteration — try five openings, keep one, edit it with references, and finish only the winner in 4K. And because the generator adapters are open, developers can build that workflow into their own tools, ComfyUI and Diffusers included, instead of being locked to one website. For short-form creators, product-video makers and small brands, that speed plus control is the real story here — not just another quality bump.

Sources: Pika official blog, "Introducing Pika H3 Flex" and "Pika H3 Flex Technical Deep Dive" (both 29 October 2026), Pika H3 Flex product page on pika.art, and the official Pika-Labs/Pika-H3-Flex pages on Hugging Face and GitHub — all opened and checked on 6 November 2026.