Pika H3 Flex: Specs, License & How to Run

Pika H3 Flex is Pika’s post-trained MiniMax H3 video-and-audio adapter family. Official hosts we opened on 2026-10-30: the ungated card Pika-Labs/Pika-H3-Flex (created 2026-10-29T17:08:44Z; last modified 2026-10-29T23:29:03Z; 166 likes / 87 downloads), the MiniMax H3 Community LICENSE, github.com/Pika-Labs/Pika-H3-Flex (178 stars / 9 forks; created 2026-10-29T17:08:59Z; last push 2026-10-29T23:29:02Z), and the launch post Introducing Pika H3 Flex (29 October 2026). API gated was false. Hugging Face pipeline_tag is image-to-video. cardData.base_model_relation is adapter on MiniMaxAI/MiniMax-H3.

This page hangs off Runway alternatives. That mapping is editorial (video-creation search). This is not a Runway clone and not a code-driven renderer. Head-to-head facts versus the other video entity already here: Pika H3 Flex vs SenseNova 6.8 Code2Video. Pixel-direct DiT already on that hub: LTX-2.5.

Open weights vs Pika-hosted extras

Copied from the card and blog — local package:

• Native output 480p / 768p, 24 fps, 32 kHz stereo mixed into the MP4
• Six-step generation on the released adapters
• Up to 12 reference files (Ref2VA: up to 9 images, 3 videos, 3 standalone audio clips)
• Durations 5–15 seconds on the model frame grid

Not in the downloadable open-weight grant (Pika-hosted API):

• 1080p / 2K / 4K upsampler
• Audio-stem separation (blog: Dialogue, Score, SFX, Foley, Ambience)
• Optional prompt enhancement
• Accelerated serving stack

The blog’s “generate a 15-second clip at 768p in as little as seven seconds of transformer computation” is a vendor DiT-compute figure, not end-to-end latency. DesignArena “#1 among open-weights / #7 overall” is a launch snapshot on the blog — self-reported, not a forkalt rank.

What we verified (2026-10-30)

Field	Value	Source
Host	huggingface.co/Pika-Labs/Pika-H3-Flex (not gated)	HF API
Created / last modified	2026-10-29T17:08:44Z / 2026-10-29T23:29:03Z	same
Likes / downloads	166 / 87	same
License (card)	other / minimax-h3-community-license-agreement	HF cardData
LICENSE file	MiniMax H3 Community License Agreement, 2 August 2026	LICENSE
Pipeline tag	image-to-video	HF API
Relation	adapter on MiniMaxAI/MiniMax-H3	HF cardData
safetensors.parameters	33,122,992,896 (~33.1B)	HF API
usedStorage	219,901,669,273 bytes (~205 GiB)	same
GitHub stars / forks	178 / 9	GitHub API
Code license	Apache-2.0 (LICENSE-CODE; CLI name pika-h3-pro)	GitHub
Base components (docs)	~210–211 GB MiniMax-H3 + ~5.5 GB adapters	card + hardware.md

runsLocally: true means the adapters and pinned base can be downloaded. The documented serve path is one GPU per request (vendor H100 80GB). That is not a laptop checkpoint.

How to run (from official README, not yet tested)

Clone Pika-Labs/Pika-H3-Flex. The CLI distribution name is pika-h3-pro. Card/GitHub name pip, Conda, or Docker. Copied example:

export MODEL_REPO=Pika-Labs/Pika-H3-Flex
export MODEL_REVISION=0ddf55ad7067ca1e76812253bd7b4600df87f759
pika-h3-pro --task t2va --prompt-file examples/prompts/t2va.txt \
  --model-repo "$MODEL_REPO" --revision "$MODEL_REVISION" \
  --resolution 480p --aspect 16:9 --duration 5 --seed 42 --output outputs/sample.mp4

Tasks named: t2va, i2va, l2va, fl2va, ref2va. Pin the revision in release.json. The card also names ComfyUI LoRAs under comfy/loras/ and Hugging Face Diffusers examples. forkalt has not run these commands.

License

LICENSE is the MiniMax H3 Community License Agreement (2 August 2026). Hugging Face license_name is minimax-h3-community-license-agreement. NOTICE on the card credits MiniMax and Mellis, Inc.

Paraphrase only, not legal advice:

• Applicable Territory is worldwide excluding the European Union, the United Kingdom, the Republic of Korea, and the United States. Use outside that territory is not authorized by this agreement (contact MiniMax for a separate license).
• Yearly revenue over USD 20 million requires prior written MiniMax authorization ([email protected]).
• Prominently display “MiniMax H3” on commercial product UIs.
• You may not use the Works or Outputs to improve any other AI model other than MiniMax H3 or its Model Derivatives.
• Example Python / CLI code is Apache-2.0; that does not cover model weights or example media.

What we did not invent

• 4K, stems, and “faster than real time” stay hosted or vendor.
• DesignArena rank stays a blog launch snapshot.
• Peak VRAM is not independently measured. GitHub 66–69 GiB is a vendor representative H100 figure.
• We do not rank this card against LTX-2.5, SenseNova 6.8 Code2Video, or Runway.

Related: LTX-2.5 (pixel DiT) — SenseNova 6.8 Code2Video (code-driven) — Pika H3 Flex vs SenseNova 6.8 Code2Video — HyperFrames (HTML→MP4) — Runway alternatives. Catalog: home. Editorial: About.