The name
Pika H3 Flex is the published family.
Pika H3 Flex is the open-weight release behind that search. Pika post-trained it on MiniMax H3. Six sampling steps write the picture and stereo audio together.
Pika H3 Flex is the published family.
Adapters on MiniMax H3, not a new network.
Local weights, ComfyUI, Diffusers, or the Pika API.
The MiniMax license excludes four territories.
Pika published Pika H3 Flex on 29 October 2026. Weights are on Hugging Face, inference code is on GitHub, and ComfyUI has workflows. The hosted path is the Pika API.
The product page is pika.art/pika-h3-flex. The README calls its playground link Try it now. That link is the API playground, not a separate local demo.
Flex runs on MiniMax H3. ComfyUI loads the adapters onto that base.
Pika’s research post describes fine-tuning, distillation from 50 steps to six, then merged reward and preference LoRAs.
Two FP32 LoRA adapters, rank and alpha 128. FL2VA uses transformer. Ref2VA uses transformer_ref.
The base files stay MiniMax’s. One workflow loads one transformer partition.
Text (t2va), first frame (i2va), last frame (l2va), or both frames (fl2va).
Up to 9 images, 3 videos, and 3 audio clips: 12 files. At least one image or video.
Six forwards. Knots: 1000, 750, 500, 250, 100, 40, 0. Generic six-step Euler is not this schedule.
32 kHz stereo, mixed into a 24 fps MP4. Local canvas is 480p or 768p.
Clone Pika-Labs/Pika-H3-Flex. Install pika-h3-pro with pip, Conda, or Docker. Python 3.12, one NVIDIA GPU.
Hugging Face Pika-Labs/Pika-H3-Flex. Both tasks are about 210 GB of base files plus 5.5 GB of adapters.
LoRAs on the MiniMax H3 base: text, reference, and first-last templates. Strength 1.0. Do not stack a turbo LoRA.
Examples in the repo, separate from the CLI, pinned to a public Diffusers commit.
Native 480p or 768p, 5 to 15 seconds, one GPU. A 5-second request is 124 frames. Checked examples use one H100 80GB.
API draft-to-video makes 1080p, 2K, and 4K. Pika’s research post says that upsampler is hosted and planned for open release.
The API returns Dialogue, Music, and Effect. The research post names five stems. Those Pika pages disagree. Use the API page.
The product page says faster than real time. The footnote is average DiT time, not end-to-end latency, and not a local measurement.
Runway, Sora, and other hosted-only video APIs keep the weights private. Flex publishes adapters. 1080p and above, stems, and prompt rewriting stay on the Pika API.
Same family. Pika describes six-step sampling with audio in the pass, against a 50-step teacher. The base license still applies.
The research post says “#1” and “up to 91%”. The product page cites another board. Those are vendor statements, not independent scores.
Local weights fit 480p or 768p where the license allows. Use the API for finishing. Read both docs before you mix them.
The adapters are Model Derivatives under the MiniMax H3 Community License. Apache-2.0 covers inference code and the Qwen3-VL encoder, not the weights.
Excluded Territories: the European Union, the United Kingdom, the Republic of Korea (South Korea), and the United States. Use in those territories is not authorized.
MiniMax says people there can request a separate license. More than 20 million US dollars in yearly revenue needs prior written authorization.
This reads the public license. It is not legal advice. The acceptable-use policy still applies inside the allowed territory.
Public pages, read in November 2026. Vendor rankings stay vendor statements. More free explainers: Knowledge Hall.