Introduction

Pika just dropped Pika H3 Flex, its first open-weights video model. It is a post-trained version of MiniMax H3 that generates video with synchronized stereo audio in six sampling steps instead of 50, accepts up to 12 reference files, and is out today in three places: the weights on Hugging Face, the inference code on GitHub, and the hosted Pika API, which adds 1080p, 2K and 4K finishing plus separate audio stems.

Pika's framing is lovely and honest: your first generation is a starting point, not the finish line. Flex is built for the tinkering loop, so you can try the weird angle, the stranger character, the idea that might be terrible, without waiting minutes per take.

What Shipped

Flex is one video-and-audio model family with two task adapters on top of MiniMax H3:

  • FL2VA handles text-to-video and endpoint-frame conditioning: text only (t2va), first frame (i2va), last frame (l2va), or both (fl2va).
  • Ref2VA handles multimodal references (ref2va): up to 9 images, 3 videos and 3 standalone audio clips, with 12 files total and at least one image or video.

Both adapters are rank-128 LoRAs on the H3 trunk, and each ships with a manifest that pins the exact six-step sampling schedule. Pika is clear that a generic six-step Euler sampler is not the same thing, so use their schedule or the official workflows.

The Numbers

  • Speed: Pika says a 15-second, 768p clip with synchronized audio can take as little as about seven seconds of transformer compute, which it calls faster than real time.
  • Resolution (local): native 480p or 768p across 11 aspect ratios, from 21:9 down to 9:21.
  • Resolution (API): 1080p, 2K and 4K through Pika's draft-to-video endpoint.
  • Duration: 5 to 15 seconds at 24 fps (a 15-second request renders 362 frames).
  • Audio: 32 kHz stereo mixed into the MP4 locally; the API can export five stems (dialogue, score, SFX, Foley and ambience).
  • Prompts: up to 8,000 characters, so long shot-by-shot direction is welcome.
  • Ranking: Pika reports #1 among open-weights models and #7 overall on DesignArena's text-to-video board, above base H3 and H3 Max.

Where to Run It Today

  • Pika API: Pika says pricing starts at $0.01 per second, and it is giving away 15 generations per day for the next two weeks on its API Club.
  • ComfyUI: Comfy-format BF16 LoRAs live in the repo's comfy/loras/ folder, with official T2V, FL2VA and R2V workflow templates and a Comfy tutorial. Load them at LoRA strength 1.0 on the matching native MiniMax H3 base.
  • Diffusers: there is an official Pika H3 Flex pipeline page plus example scripts for first-frame and reference generation.
  • Local CLI: clone the repo and use the pika-h3-pro command, which pulls the pinned base parts and the adapter you pick.

Hardware is the honest catch. The bundled H3 base parts are about 210 GB on disk (a workflow loads only the trunk it needs), and Pika's tested CLI runs were on an H100 80GB, with --cpu-offload for big requests. This is a workstation or cloud-GPU model, not a laptop toy.

Grab the ComfyUI Adapters

python -m pip install huggingface_hub
export COMFY_REVISION=$(python -c 'from huggingface_hub import HfApi; print(HfApi().model_info("Pika-Labs/Pika-H3-Flex").sha)')
hf download Pika-Labs/Pika-H3-Flex --revision "$COMFY_REVISION" --include 'comfy/*' LICENSE NOTICE 'licenses/*' --local-dir ./pika-h3-flex-comfy

Then drop the two files from comfy/loras/ into your ComfyUI models/loras/ folder. Pika notes that it checked the tensor conversion but has not yet verified a full end-to-end ComfyUI render, so treat day-one Comfy output as an early ride.

Prompting Tips From the Showcases

Flex rewards direction. Pika's showcase prompts read like shot lists: timestamped camera moves (an overhead open, a crane down, a whip pan, an orbit), explicit "normal real-time speed, no slow motion" notes, and a separate Sound: line for diegetic effects and music. For Ref2VA, name what each reference contributes, for example the character from one image, the world from another, and only the motion from a video.

Here is a paste-ready starter in that style:

Image 1 is the character: keep her face, outfit and colours exactly as in Image 1. A rooftop garden at blue hour above a neon city. 0-4 s: wide shot, slow orbit as she waters glowing lanterns that hang like fruit from the vines. 4-8 s: push in to a waist-up close-up as one lantern floats free and drifts past her face; she laughs. 8-12 s: crane up and back to reveal the whole city lighting up in lilac and aqua. 12-15 s: hold on the wide view as the last lantern joins the skyline. Normal real-time speed, no slow motion. Soft 35mm film look. Sound: gentle wind, a watering can trickle, distant city hum, a warm lo-fi piano melody.

License Notes

The adapters are model derivatives of MiniMax H3, so they follow the MiniMax H3 Community License Agreement, not a plain permissive license. The example Python code is Apache 2.0, and the Qwen encoder parts carry their own Apache 2.0 terms. Read the license and NOTICE files before you build a product on it.

Original Source

Model card: Pika-Labs/Pika-H3-Flex on Hugging Face. Launch post: Introducing Pika H3 Flex.

Conclusion

Six steps, twelve references, sound baked in, and open weights from day one: H3 Flex is the kind of release that turns a slow render queue into a sketchbook. If you have the GPU, the ComfyUI templates make it a fun weekend experiment; if you do not, the API free tier is a gentle way to test the iteration loop first.

Want to play with the H3 family without a 210 GB download? Try this on Gen → https://artrealmai.com/gen?utm_source=magazine&utm_campaign=gen&utm_content=pika-h3-flex-open-weights-six-step-video-audio-12-references

—Aurelia ♡