Pika Labs releases open video model H3 Flex: 6 steps instead of 50

Pika Labs has released H3 Flex, an open model for video generation. It was fine-tuned from MiniMax H3, and the rendering process dropped from 50 steps to 6. A step here is one pass of a large neural network (a diffusion transformer, or DiT) that turns random noise into an image by removing the noise gradually. The fewer the passes, the faster and cheaper the result. Pika says that denoising a 15-second clip at 768p takes 7 seconds, and that the model as a whole runs 1.8 times faster and 80% cheaper than H3 Max.

Why there are 8x fewer steps, but the speedup is only 1.8x

Video models are expensive precisely because of the step count. Each pass processes dozens of frames at once, so 50 iterations add up to a noticeable GPU bill. Pika took MiniMax H3 and fine-tuned it with a preference-based method. The model receives pairs of outputs in which one is better than the other, and learns to produce the good one quickly, without relying on a long chain of refinements. This technique is often used to cut the number of steps without an obvious loss in quality.

A note of caution: the step count fell by a factor of 8.3, but the overall speedup relative to H3 Max is 1.8x. The 7-second figure applies to denoising only. The full pipeline has other stages: parsing the prompt and references, decoding frames, audio, upscaling. These most likely keep the speedup from scaling proportionally. The speed and price figures come from Pika itself, and the article cites no independent measurements.

Separate audio tracks and 12 references

The main practical difference of H3 Flex is not only speed but also what it hands to the editor. The model exports five independent audio tracks:

• dialogue;
• soundtrack;
• sound effects;
• noise;
• ambient (background sound of the environment).

Editors don't have to pull a voice out of a single mixed track: they can remove the music, rewrite the lines, and keep the street noise. For advertising and short clips, this simplifies post-production considerably.

The model accepts up to 12 references as input. These can include sketches, audio samples, and moodboards. That lets you define a character's look, a style, or a sound without trying to describe everything in a single text prompt.

The model also offers:

• draft generation: quick trial versions to help you choose a direction;
• targeted editing of characters or camera movement without regenerating the whole clip;
• upscaling to 4K.

Drafts pair well with cheap, fast inference: you can run many trials and launch the expensive final render only for the one that works.

Pricing, open weights, and what to check before adopting it

Pika has published the inference code and model weights openly (a GitHub repository and a Hugging Face page) and launched an API. Generation costs:

• from 1 cent per second of video in text-to-video mode;
• up to 3 cents per second when the clip is generated from references.

By simple arithmetic, a 15-second clip will cost roughly 15 cents when generated from text and up to 45 cents when working with references. This is a rough guide; the final bill depends on resolution and settings. The model will be built into the Pika.art platform on November 2.

On quality, H3 Flex ranks eighth in the overall DesignArena ranking of video models. That is not leadership, but for a fast, inexpensive open model the result is notable. Eight competitors rank higher, and the gap to the leaders is worth evaluating on your own tasks.

What this means for practitioners. Studios and developers of video services now have an open model with separate audio tracks and a reasonable price. It can be deployed on your own GPUs or accessed through the API. Before commercial use, read the license in the repository and run typical scenarios yourself: a 7-second denoising time does not guarantee the same time for the whole pipeline. For the market, this is another signal: the video-model race is shifting from how good a frame looks to the price per second and the convenience of editing.