WaveSpeedAI
← Blog/MiniMax H3

Pika H3 Flex Explained: What It Is and How to Try the H3 Family via API

Pika H3 Flex is a six-step open-weights video adapter built on MiniMax H3. What it changes, what to check in its license, and how to call H3 text, image, and reference-to-video endpoints today.

By WaveSpeedAI5 min read
#pika-h3-flex#minimax-h3#video-api#open-weights#ai-video

Pika released Pika H3 Flex on October 29, 2026. It is easy to misread as a new base model. It is not: the open release is a pair of adapters that sit on top of MiniMax H3 and generate video with synchronized audio in six sampling steps.

This post separates what Pika published from what is still a vendor claim, then shows how to run the H3 family through an API today. To be clear up front: WaveSpeedAI does not list Pika H3 Flex itself. The endpoints below are the MiniMax H3 models.

What Pika H3 Flex actually is

According to Pika’s model card, Flex is a post-trained version of MiniMax H3 delivered as two LoRA adapters:

  • FL2VA for text-to-video and first/last-frame conditioning.
  • Ref2VA for multimodal references.

The published specifications:

ItemPika H3 Flex
SamplingSix steps
Native resolution480p and 768p
Duration5–15 seconds, 24 fps
Audio32 kHz stereo, generated with the video
Aspect ratios11, from 21:9 to 9:21
ReferencesUp to 12 files: 9 images, 3 videos, 3 audio clips

Higher resolutions (1080p, 2K, 4K), audio stems, and prompt enhancement are separate hosted services on Pika’s own API, not part of the downloadable weights.

What is verified and what is a claim

Pika’s technical write-up reports small margins over base MiniMax H3 on its own metrics, and says the model is distilled from a 50-step setup to six. Those are vendor-run numbers. Launch coverage also cites a DesignArena text-to-video ranking of #7 with an Elo of 1293, against 1277 for base H3, as reported by Kie. That is a 16-point gap, not a large one.

I found no independent side-by-side tests against other fast H3 variants. If you care about a specific shot type, run your own prompts rather than relying on a leaderboard.

Check the license before you self-host

The adapters are Model Derivatives of MiniMax H3 and are distributed under the MiniMax H3 Community License. The license text defines its applicable territory as worldwide excluding the European Union, the United Kingdom, South Korea, and the United States. It also covers hosted services and commercial use. Read it with your legal team before building a product on the weights.

Hardware matters too: Pika’s package runs one request on one GPU, and its tested 480p and 768p requests used an H100 with 80 GB. The base components alone are about 210 GB.

Try the H3 family through an API

If you want to evaluate H3-family video and audio generation without provisioning GPUs or downloading weights, WaveSpeedAI exposes three MiniMax H3 endpoints:

EndpointInputPage
minimax/h3/text-to-videoPromptText to video
minimax/h3/image-to-videoFirst frame, optional last frameImage to video
minimax/h3/reference-to-videoUp to 9 images, 3 videos, 3 audio filesReference to video

Current published parameters and prices, as listed on the model pages on October 31, 2026:

ResolutionPrice per second5 seconds15 seconds
768p$0.10$0.50$1.50
2K$0.14$0.70$2.10
  • Duration is 4–15 seconds as an integer.
  • Text-to-video accepts 21:9, 16:9, 4:3, 1:1, 3:4, and 9:16.
  • For reference-to-video, the first five reference images are free; each additional image costs $0.05. Audio cannot be sent alone, and at least one image or video is required.
  • The model pages state that the final task charge prevails, so check the page before budgeting a large batch.

Submit a task and poll for the result

import os
import time
import requests

headers = {"Authorization": f"Bearer {os.environ['WAVESPEED_API_KEY']}"}

submit = requests.post(
    "https://api.wavespeed.ai/api/v3/minimax/h3/text-to-video",
    headers=headers,
    json={
        "prompt": "A street musician plays violin in a rainy European alley, "
        "neon signs reflecting on wet cobblestones, slow dolly-in",
        "aspect_ratio": "16:9",
        "resolution": "768p",
        "duration": 5,
    },
    timeout=30,
)
submit.raise_for_status()
task_id = submit.json()["data"]["id"]

while True:
    result = requests.get(
        f"https://api.wavespeed.ai/api/v3/predictions/{task_id}/result",
        headers=headers,
        timeout=30,
    )
    result.raise_for_status()
    data = result.json()["data"]
    if data["status"] == "completed":
        print(data["outputs"])
        break
    if data["status"] in {"failed", "cancelled", "timeout", "deleted"}:
        raise RuntimeError(data.get("error") or data["status"])
    time.sleep(2)

For retries, polling intervals, and result handling, see the API getting-started guide and the MiniMax H3 API guide.

Which route fits which job

  • Try an H3-family model quickly or ship a feature without GPU operations: use the API above.
  • Need local control, pinned weights, or fine-tuning: use Pika’s or MiniMax’s open weights, after the license review and a GPU budget.
  • Need a hosted fast-iteration loop from a specific vendor: compare vendors on wall-clock time, not only on step count; see H3 Max vs MiniMax H3.
  • Cost planning for H3 video: see MiniMax H3 pricing per second.

A six-step model changes the cost of iteration; it does not change the need to test on your own prompts. Pick two or three representative shots, run them through each route you are considering, and compare usable outputs per dollar rather than a single best clip.

Share