Pika Launches Pika H3 Flex, a Video Model Built for Professional Iteration

The highest-ranked open weights video model on DesignArena’s text-to-video board, post-trained on MiniMax H3

SAN FRANCISCO--(BUSINESS WIRE)--Pika, the Lightspeed-backed gen media darling, is on a mission to make AI support the way creative professionals actually work. In the last 3 months, they’ve launched a creation product that obscures models in favor of function, a gen media API platform that obliterates the 3x industry markup, and an official ChatGPT plugin that turns the chat into an immersive creative studio.

Now, they’re launching a video model that’s built for professional iteration: Pika H3 Flex. Pika’s research team post-trained the model on MiniMax H3 to be faster and more controllable, while maintaining the quality professionals need.

The result: Pika H3 Flex generates faster than real-time, offers resolutions up to 4K, and exports up to 5 separate audio stems. And it’s ranked #7 in video quality, above base H3, H3 Max, and Seedance 2.5 on DesignArena’s text to video board.

“Most video models are built on the assumption that you plan to finalize your first generation. We know that professionals will always tinker, tweak, and refine. We built Pika H3 Flex with that in mind,” said Chenlin Meng, Chief Scientist and Co-founder at Pika.

In the spirit of creative iteration, Pika is also open-sourcing its weights.

A model that hits high quality marks

DesignArena’s independent results mirror Pika’s own evaluations. In Pika’s automated benchmarks, Pika H3 Flex won the overwhelming majority of comparisons against MiniMax H3, fal H3 Max, and HeyGen Video’s quality model across all four video-quality metrics tested. In Pika’s separate human preference evaluation, Flex ranked highest among the models tested for human anatomy, complex composition, audio sync, visual style, and camera control.

Built to keep up with the creative process

Pika developed H3 Flex to make creative iteration faster while maintaining the quality and control professionals need. Starting with MiniMax H3, the team fine-tuned the model on curated video, audio, and reference-based tasks spanning camera movement, character consistency, motion transfer, and editing.

“Our goal was to give creators more room to experiment without asking them to compromise on quality,” said Cade Li, Research Lead at Pika. “We brought generation down from 50 steps to just six, then refined the faster model for visual quality and creative control. That means creators can spend less time waiting and more time getting the shot right.”

The team used reward feedback and preference judgments calibrated against human evaluations to refine the accelerated model, preserving, and on several measured quality dimensions improving upon, the quality of its 50-step teacher.

Pika also optimized the system that runs the model around its six-step process. Together, these advances allow Flex to generate a 15-second video with synchronized audio at 768p in as little as seven seconds of DiT denoising time, giving creators more room to explore and revise before finishing their chosen output at 1080p, 2K, or 4K.

Model highlights

Faster than real-time: Pika H3 Flex can generate a 15 second clip in as little as 7 seconds, for creation at the speed of thought.*

Separate audio stems: export up to 5 audio stems to isolate and refine any facet of the soundscape: Dialogue, Score, SFX, Foley, and Ambience.

Draft mode: start at 768p and upscale to 4K in finishing mode when you’re ready.

Multi-modal creation: create from up to 12 references, including sketch, voice sample, image, sound, video, and moodboard.

Priced to play: 1.8x faster than fal H3 Max, and up to 80% cheaper.

Precision editing: swap characters, camera moves, looks, voices, and more.

Open weights: iterate on the model itself.

The stats

On DesignArena’s text-to-video board, Pika H3 Flex is the highest-ranked open weights video model, and ranks #7 in video quality, above base H3, H3 Max, and Seedance 2.5.

Generates 15-second clips at 768p in as little as 7 seconds.*

1.8x faster than H3 Max, on average, across matched t2v, i2v, and r2v comparisons.

*DiT denoising time, not guaranteed end-to-end API latency.

Where to access Pika H3 Flex

Pika API Club: pika.art/pika-h3-flex

Open weights: available on GitHub and Hugging Face

Pika.art: available on the creation product starting November 2nd

For the first two weeks after launch, Pika is offering members and subscribers 15 free generations per day, and all generations at 50% off.

About Pika

Pika is an AI video company backed by Lightspeed, known for launching some of the most widely used generative video tools since its founding in 2023, including Pikaffects. Pika H3 Flex is the company’s post-trained video model built for professional iteration. For more information, visit pika.art.