Pika Cuts MiniMax H3 From 50 Denoising Steps to Six and Posts the Video-and-Audio Adapters on Hugging Face

Pika has put the weights behind its newest video model on Hugging Face, and the headline number is six. Pika H3 Flex, announced on October 29, 2026, generates a video clip with a matching stereo soundtrack in six sampling steps, down from the 50 its starting model uses. The weights live in the Pika-Labs/Pika-H3-Flex repository, with inference code on GitHub and support in ComfyUI and Hugging Face Diffusers.

Pika’s research post calls Flex “an open-source omni video model family” and says it ranks first among open-source text-to-video models. That ranking and the comparisons behind it come from Pika’s own evaluation.

Built on MiniMax H3, shipped as adapters

Flex is not a model trained from scratch. Pika post-trained MiniMax H3, an open-weight audio-video model from MiniMax, and released the result as two LoRA adapters at rank and alpha 128. The FL2VA adapter handles text prompts and first- or last-frame conditioning. The Ref2VA adapter handles multimodal references.

The repository also bundles the MiniMax H3 base components in Diffusers format, about 210 GB in total, though a given workflow loads only the transformer partition it needs. Hugging Face lists the model at 33B parameters.

Licensing has two layers. The adapters are described as “Model Derivatives of MiniMax H3” under the MiniMax H3 Community License Agreement, and the bundled base components keep that agreement. The Qwen text encoder and the example Python code are Apache 2.0, but the card says the code license “does not cover model weights or example media.” Commercial users need to read the MiniMax terms rather than assume a permissive license.

Three stages of post-training

Pika describes three stages. First, supervised fine-tuning on an in-house audio-video dataset curated for aesthetic quality, which kept the 50-step schedule and became the teacher. Second, step distillation: standard Distribution Matching Distillation was unstable in Pika’s experiments, so it used Projected Distribution Matching Distillation, which produced a six-step student. Third, reward learning on the distilled model, with Reward Feedback Learning for aesthetics and clarity and Direct Preference Optimization for instruction following and realism, trained as separate LoRAs and merged into one adapter at no extra inference cost.

The model card warns that the six-step schedule is specific to the release. “Generic six-step Euler sampling is not equivalent,” it says, and each manifest pins the base revision, adapter checksum and exact schedule.

What runs locally

Local inference produces native 480p or 768p video at 24 frames per second across 11 aspect ratios, from 21:9 to 9:21, with 32 kHz stereo audio mixed into an MP4. Requests run from 5 to 15 seconds, rounded up to the model’s frame grid, so a 5-second request yields 124 frames. Prompts can be up to 8,000 characters.

Ref2VA accepts up to nine images, three videos and three standalone audio clips, with 12 reference files in total. The card describes these as input limits, not a promise that every maximum-size combination fits in GPU memory. The CLI runs one request on one GPU, and representative 480p and 768p requests passed on an H100 80GB, with CPU offload for larger jobs.

Some features stay on Pika’s hosted API: 1080p, 2K and 4K output from a 1B-parameter upsampler, five separate audio stems for dialogue, score, effects, foley and ambience, and prompt enhancement. Pika says it will open-source the upsampler, without a date.

The numbers, and their limits

On Pika’s 768p text- and image-to-video test, Flex scored highest on all five visual and audio quality metrics against MiniMax H3, Fal H3 Max and HeyGen Video. For example, it scored 12.3 on DOVER++ technical quality against 11.7 for MiniMax H3, and 7.02 on audio aesthetics against 6.85. Audio-video and text-video alignment were roughly level across the four systems, and Fal H3 Max scored slightly higher on audio-video alignment.

Pika’s blind human preference study put Flex ahead on anatomy, complex composition and audio sync for text and image prompts, and on style and camera motion for reference prompts. Pika also reports where it lost: other systems were preferred on physics, multi-scene narrative, lighting and materials, and action-heavy reference tasks.

For speed, Pika says inference optimizations cut full generation of a five-second 768p clip from 3.409 to 2.129 seconds, measured on eight B300 GPUs. That is Pika’s serving stack, not the single-GPU package in the repository.

Every comparison in the post was run by Pika, and the showcase clips are labeled “not release-validation renders or throughput benchmarks.” The weights are public, so independent testing can now check the claims.