On Oct. 29, 2026, Pika introduced Pika H3 Flex, a post-trained version of MiniMax H3 aimed at professional iteration: faster drafts, richer reference control, and synchronized stereo audio in the same generation pass. Pika is releasing open weights for the model's generator adapters and inference code, while keeping upscaling, stem separation, and prompt enhancement on its hosted API. Flex is also scheduled to arrive in the Pika creation product on November 2.
Key facts
| Item | Detail |
|---|---|
| Model | Pika H3 Flex — post-trained MiniMax H3 omni video-and-audio family (Pika) |
| Sampling | Six generator steps vs. 50 for the teacher (Pika: “six steps instead of 50”) |
| References | Up to 12 files via Ref2VA (up to 9 images, 3 videos, 3 standalone audio clips) |
| Native output | 480p or 768p, 24 fps, 5–15 second clips, 11 aspect ratios, 32 kHz stereo MP4 (H.264 + AAC) |
| Hosted API extras | 1080p, 2K, and 4K upscaling; five audio stems; prompt enhancement |
| Rankings (launch snapshot) | #1 among open-weights models and #7 overall on DesignArena's text-to-video leaderboard (Pika, at the time of writing) |
| Open release | Weights on Hugging Face (Pika-Labs/Pika-H3-Flex); code on GitHub; ComfyUI and Diffusers paths |
| Licence | MiniMax H3 Community License Agreement (adapters are Model Derivatives of MiniMax H3) |
What Pika released
In Introducing Pika H3 Flex, Pika describes Flex as its post-trained MiniMax H3 model, built so the first generation is a stronger starting point and later versions stay under creative control. Pika said it post-trained MiniMax H3 to generate in six steps instead of 50, then refined the faster model for quality. At the time of writing, Flex ranks #1 among open-weights models and #7 overall on DesignArena's text-to-video leaderboard; Pika notes that position reflects the launch snapshot.
The deep dive outlines two adapters on one video-and-audio family: FL2VA for text and endpoint frames and Ref2VA for multimodal references, with generator weights and inference code released for local use under licence terms.
What it can do
Flex outputs video with synchronized stereo audio. The Pika API can return five separate audio stems—Dialogue, Score, SFX, Foley, and Ambience—for mix work without re-rendering picture.
Pika supports up to 12 reference files in one generation. The model card caps Ref2VA at 9 images, 3 videos, and 3 standalone audio clips (at least one image or video required). FL2VA adds text-to-video, first-frame, last-frame, and first-and-last-frame modes. Pika's examples stress directed camera moves, multi-shot cuts, dialogue, and reference-guided edits such as character or setting swaps while preserving performance.
Local inference natively delivers 480p or 768p, 24 fps, 5–15 second clips, 11 aspect ratios (21:9 through 9:21), and 32 kHz stereo in MP4 (H.264 + AAC). Draft at 768p, then upscale favorites through Pika's hosted upsampler.
Speed claims
Pika says Flex can generate a 15-second clip with synchronized audio at 768p in as little as seven seconds of transformer computation—and stresses that seven seconds measures transformer (DiT) computation for a 15-second 768p clip in a controlled benchmark, not total end-to-end API latency. Reference inputs, finishing, stems, hardware, and service overhead all change real wait times.
In training, Pika distilled a 50-step teacher to a student that generates joint audio-video in six denoising steps—about 8× fewer than the teacher in its write-up.
What is open vs. API-only
Open weights cover adapters and inference for text, endpoint-frame, and reference-to-video modes at 480p/768p with synced audio—on Hugging Face and GitHub, plus ComfyUI and Diffusers paths. Hosted-only pieces include 1080p, 2K, and 4K upscaling, audio-stem separation, prompt enhancement, and Pika's accelerated serving; local performance depends on your hardware and setup, Pika warns.
Pika's Table 1 cites a starting cost of $0.01 per output second for Flex on its API comparison—attribute that to Pika's pricing table, not self-hosting economics.
Hardware and licence
The package runs one request on one GPU; Pika reports representative 480p/768p CLI runs on an H100 80GB, with --cpu-offload for large jobs. Base components total roughly 210 GB on disk. Adapters ship under the MiniMax H3 Community License Agreement as Model Derivatives of MiniMax H3.
Why it matters
Analysis. Six-step omni-video with twelve reference slots targets pro iteration—many drafts, selective fixes, then a finish pass—rather than single-shot demos. Splitting open weights from 4K upscaling and stems is the same playbook as other frontier media APIs.
Longer rival timelines such as Kling 4.0 Goes Global: 30-Second AI Video and 10 Keyframes push duration; native audio aligns Flex with conversations sparked by DeepMind's Genie 4 World Model Adds Native Audio, even though Flex is a production generator, not an interactive world model. Pika's deep dive reports up to 91% gains over leading competitors in its own human preference evaluations—useful context, not third-party proof.
What to watch
Flex reaches the Pika creation product on November 2. Track whether DesignArena's launch snapshot holds, how sub-H100 users cope with memory limits, and whether teams adopt Ref2VA-heavy pipelines or stay on FL2VA endpoint control before paying for 4K finishes.