Pika H3 Flex

Pika's six-step post-train of MiniMax H3 makes video with sound from text, frames or 12 references, but keeps H3's territory limits.

In short. Pika H3 Flex is an add-on for the open MiniMax H3 video model, made by the video company Pika. It makes a clip of 5 to 15 seconds with matching stereo sound in six sampling steps instead of 50, from text, first and last frames, or up to 12 reference images, videos and audio clips. You still need the large MiniMax H3 base files, and the license keeps H3's limits: it does not cover the United States, the European Union, the United Kingdom or South Korea.

Summary

Pika H3 Flex is a post-trained version of MiniMax H3 that Pika published on 29 October 2026. What you download is a pair of LoRA adapters, each about 2.8 GB in FP32 at rank 128, for one video-and-audio model: FL2VA covers text and first-frame or last-frame conditioning, and Ref2VA covers references. Ref2VA takes up to 9 images, 3 videos and 3 audio clips, with at most 12 files in total. The repository also carries the MiniMax H3 base components in Diffusers format, which Hugging Face contributed, about 211 GB in all; one workflow loads only its own transformer, so a single setup is about 147 GB with the adapter. The adapters are Model Derivatives of H3 and use the MiniMax H3 Community License, while the Qwen3-VL text encoder is Apache 2.0.

Pika describes three post-training stages. It first fine-tuned H3 on an in-house, aesthetically curated audio-video dataset, then used projected distribution matching distillation to cut sampling from 50 steps to six, and finally trained reward-feedback and preference-optimization adapters that were merged into one. The open package makes clips of 5 to 15 seconds at 24 frames per second, with 32 kHz stereo audio, at native 480p or 768p across 11 aspect ratios, from prompts of up to 8,000 characters. Pika's blog also mentions 1080p, 2K and 4K output, separate audio stems and prompt enhancement, but those are hosted API features and are not in the download.

Pika says it ranks first among open-source text-to-video models. In its own tests against the official H3, Fal H3 Max and HeyGen Video at 768p, it has the highest score on five visual and audio quality metrics and leads human preference on anatomy, complex composition and audio sync for text and image input, and on visual style and camera motion for references; other systems were still preferred for physics, multi-scene stories, lighting and action-heavy references. Pika says transformer denoising time stays below the clip's length, and measures 2.129 seconds of full inference for a five-second 768p clip on eight B300 GPUs. The open package is different: it runs each request on one GPU and was tested on an H100 with 80 GB, where a default 480p, 5-second request allocated about 66 to 69 GiB. A 768p, 15-second request needs CPU offload, which used about 29 to 32 GiB on the GPU and about 94 GiB of host memory, and Pika says these are not minimum requirements. H200 and B300 have not been qualified for the package.

The license is H3's, so the territory limits come with it: the grant excludes the European Union, the United Kingdom, South Korea and the United States, and products earning more than US$20 million a year need MiniMax's written authorization. ComfyUI workflows, Comfy-format LoRAs and Diffusers examples are included.

Good for

  • Short clips with synchronized speech and sound made locally, faster than the 50-step H3 base, if you are inside the license's territory
  • Multi-shot storytelling and camera-directed prompts, written as timed shots, which Pika lists first
  • Reference-guided animation, motion graphics and product films, with up to 12 reference files
  • ComfyUI and Diffusers pipelines, since workflows and loading examples are included

Watch out for

  • The license is the MiniMax H3 Community License, which excludes the European Union, the United Kingdom, South Korea and the United States, and its revenue and attribution conditions
  • The "real-time" speed is measured on eight B300 GPUs, while the open package runs one request on one GPU and was tested only on an H100 with 80 GB
  • 1080p, 2K and 4K output, audio stems and prompt enhancement are hosted API features; local output is 480p or 768p
  • You need the H3 base components, about 211 GB for both transformers plus the encoder and autoencoders, so one setup uses about 147 GB of disk
  • All quality, preference and speed results are Pika's own, against H3, Fal H3 Max and HeyGen Video

Performance claims come from the developer and have not been independently verified.

How much memory does it need?

Memory figures for each variant of Pika H3 Flex. Image, video, speech and music pipelines also load a text encoder and a decoder, so the memory needed can be larger than the main checkpoint.

Variants Parameters Download size Memory needed Fits on
FL2VA adapter (text and endpoint frames) on the MiniMax H3 base BF16 33.1B 147 GB 169 GB Big-memory desktop
Ref2VA adapter (multimodal references) on the MiniMax H3 base BF16 33.1B 147 GB 169 GB Big-memory desktop

Download size covers every component needed to run the model, such as the generator, text encoder and decoder. Memory needed is a stated or published measured minimum where one exists; otherwise it is the download size plus about 15%. Larger images, longer clips and higher resolutions need more.

The license in plain words

A community license that lets you use, modify and distribute MiniMax H3, but only in the "Applicable Territory": the whole world except the European Union, the United Kingdom, South Korea and the United States. Commercial products earning over US$20 million a year need MiniMax's written authorization, every commercial product must show "MiniMax H3" in its interface, and you may not use the model or its outputs to train other AI models. Get legal advice, especially if you or your users are in an excluded region.

  • Attribution or model name must be shown
  • Free commercial use only below a revenue cap
  • Not licensed in some regions
  • Acceptable-use policy applies
  • Outputs must be filtered or reviewed, and AI generation disclosed

More about MiniMax H3 Community

See also

Frequently asked questions

What is Pika H3 Flex?
Pika H3 Flex is an add-on for the open MiniMax H3 video model, made by the video company Pika. It makes a clip of 5 to 15 seconds with matching stereo sound in six sampling steps instead of 50, from text, first and last frames, or up to 12 reference images, videos and audio clips. You still need the large MiniMax H3 base files, and the license keeps H3's limits: it does not cover the United States, the European Union, the United Kingdom or South Korea.
Can I use Pika H3 Flex commercially?
Yes, with conditions. Pika H3 Flex is released under the MiniMax H3 Community, which allows commercial use but adds extra rules: Attribution or model name must be shown; Free commercial use only below a revenue cap; Not licensed in some regions; Acceptable-use policy applies; Outputs must be filtered or reviewed, and AI generation disclosed. Read the license text before you ship a product.
Is Pika H3 Flex open source?
Pika H3 Flex is open-weight, not open source in the strict sense. You can download the model files, but the MiniMax H3 Community adds conditions that the Open Source Initiative definition does not allow.
Can I run Pika H3 Flex on my own computer?
Only on a machine with a lot of memory. The smallest variant, FL2VA adapter (text and endpoint frames) on the MiniMax H3 base, needs about 169 GB at 4-bit precision. Desktops with 128 to 512 GB of unified memory can hold it, at modest speed.
When was Pika H3 Flex released?
Pika H3 Flex was released by Pika on October 29, 2026.

Official links

Sources

Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.

Release date and facts last verified November 11, 2026.

Related models

Model Released Parameters Context License Smallest variant needs
MiniMax H3 MiniMax 33.1B – MiniMax H3 Community (Conditional) 166 GB Big-memory desktop
LTX-2.5 LTX (Lightricks) 22B – LTX-2.x Community (Conditional) 16 GB Laptop or phone
Wan 2.2 Alibaba Wan 5B – 27B 14B active – Apache 2.0 (Permissive) 24 GB Gaming GPU
MiniMax H3-Image MiniMax 33.1B – MiniMax H3-Image Community (Conditional) 155 GB Big-memory desktop