15 seconds of video in 7 seconds of computation: Pika H3 Flex accelerates AI creation

Pika H3 Flex promises faster video generation, multiple references, output up to 4K, and separate audio tracks, with open weights.

Producing a 15-second video with synchronized audio in just seven seconds of computation: the performance claimed by Pika illustrates the progress made in accelerating video generation models. With H3 Flex, the company introduces a new version derived from MiniMax's H3 model, specialized in rapid iterations and creative adjustments. It combines faster-than-real-time generation in certain configurations, the use of multiple references, targeted editing, and the separation of audio elements. A portion of this technology is also distributed with open weights, allowing developers to run it on their own infrastructures.

Launched on October 29, 2026, H3 Flex entered directly into seventh place in the overall DesignArena leaderboard for text-to-video generation, according to results presented by Pika. It also occupies the top position among open-weight models on this leaderboard, ahead of MiniMax H3, H3Max, and Seedance 2.5. Six steps instead of fifty to produce a video

The most significant improvement lies in the generation process. Rather than developing an entirely new, independent model, Pika's teams conducted additional training on MiniMax H3 to reduce the number of steps required for video synthesis.

In its technical article on H3 Flex, the company explains moving from 50 steps to just six by combining model specialization methods and reward-based optimization. The goal of the result is to maintain sufficient visual quality while significantly reducing the amount of computational power required.

According to Pika, H3 Flex runs an average of 1.8 times faster than H3Max on comparable text-to-video, image-to-video, and reference-guided video generation setups. The company also announces cost reductions of up to 80%, depending on the modes used.

The most dramatic demonstration involves a 15-second 768p clip with synchronized audio, which can be computed in seven seconds.

A clarification is necessary, however: this result corresponds to the Transformer's computation time in a controlled environment, not the total delay between sending a prompt and receiving the file. Processing references, finishing operations, server availability, and transfers can increase the time the user actually waits. Up to twelve references to guide a creation

Beyond speed, H3 Flex aims to address a recurring difficulty in generative video tools: maintaining control over scene elements across modifications.

The model accepts up to twelve reference files during a single generation, combining images, videos, and audio clips. Developers can notably provide it with a product photograph, a character board, a sketch, a visual moodboard, or a motion clip.

The system then allows users to specify, via text prompts, the role of each reference in the final output. This can include preserving a character's appearance, reusing a camera movement, or modifying the environment without replacing the entire clip.

The reference-guided editing mode also supports targeted interventions on existing videos. Pika specifically highlights replacing an outfit, an accessory, a character, or a background, while attempting to preserve the movement and pacing of the original shot.

These features aim to reduce the number of generations required to reach a satisfactory result. Above all, they enable a shift from a process where each new instruction risks rebuilding the entire scene to a workflow of successive adjustments around an initial proposal. Audio separation becomes a post-production tool

One of the most interesting new features for professionals involves the soundtrack. H3 Flex can generate videos accompanied by synchronized audio, but Pika also offers the option to export up to five distinct audio tracks.

These tracks correspond to dialogue, music, sound effects, foley, and ambience. They are provided in WAV format so they can be imported into editing or mixing software.

This separation allows users, for example, to lower the music without altering conversational volume, replace a sound effect, or tweak an ambient track independently of the rest of the composition.

Until now, videos generated with their native soundtrack were often delivered with a pre-mixed audio track, complicating precise post-production edits. Access to individual audio components brings the generative process closer to traditional audiovisual production workflows.

However, a distinction must be made between this feature and the freely distributed version. Audio track separation remains a service provided through Pika's hosted infrastructure and is not included in the open-source generation components released for local execution. True 4K output, but no native generation

H3 Flex offers 5 to 15-second clips at 24 frames per second across various aspect ratios ranging from 21:9 to 9:21.

The model's native resolutions are limited to 480p and 768p. To achieve 1080p, 2K, or 4K output, Pika utilizes an upscaling system that operates after the initial generation.

The benefit of this approach is the ability to work on a lightweight version first and upscale only the selected output. The system retains the video sequence and its audio when upgrading to a higher resolution, avoiding the need to run an entirely new generation.

Pika even offers a billing model tailored to this method: when a creation initially generated in low resolution is finalized in 4K, the user only pays the price difference between the two tiers.

This distinction separates the product's actual output capabilities from native high-resolution generation. The final output can be delivered in 4K, but the generative model does not natively render at this resolution. Open weights, with features that remain proprietary

The launch is accompanied by the release of the weights on Hugging Face and the inference code on GitHub. Developers can use the model's adapters, explore its inner workings, and build their own integrations under the terms of its license.

This distribution covers text-to-video generation, creation from initial or final