The Pulse
Pika Releases H3 Flex, an Open-Weight Video-and-Audio Model
Pika released H3 Flex, a video-and-audio model post-trained on MiniMax H3, with generator adapters and inference code for local use. Its open-weight release supports synchronized audio at 480p and 768p, while 4K finishing and separate audio
AI.info Team · Written with AI · fact-checked ·
“Your first generation is a starting point. The creative work is what happens next.”
The Pika Team
Pika announced H3 Flex on October 29, describing it as a post-trained version of MiniMax H3 that creates video with synchronized audio. The company is releasing generator adapters and inference code so developers and researchers can run and customize the model on their own hardware. The launch arrives with an important distinction: the downloadable release does not include every feature available through Pika’s hosted service.
What H3 Flex puts in the open-weight release
The released code and adapters cover text-to-video, generation from first and last frames, and reference-to-video using multiple kinds of input. Local generation produces synchronized audio at 480p or 768p; the launch post says Flex can accept up to 12 reference files, combining images, video and audio. Pika’s technical account describes the model as a six-step generator distilled from a 50-step version of MiniMax H3.
Pika says a 15-second clip at 768p can take as little as seven seconds of transformer computation. That figure measures the model’s computation in a controlled benchmark, not the time a user necessarily waits for a completed API request. Reference processing, hardware, service overhead, audio-stem separation and high-resolution finishing can add time.
4K finishing and audio stems stay hosted
The open release does not include the upsampler for 1080p, 2K and 4K output, the audio-stem separation system, optional prompt enhancement or Pika’s accelerated serving stack. Those remain Pika-hosted features, and local performance depends on the user’s hardware and setup. Flex generates natively at 480p or 768p; Pika’s hosted upsampler raises selected clips to higher resolutions.
Hosted API users can request five separate audio stems: dialogue, score, sound effects, foley and ambience. The separation gives editors individual sound components to adjust rather than one combined soundtrack. Pika says the base model generates video with synchronized stereo audio, while stem separation is a separate hosted capability.
Pika’s speed and quality claims
Pika reports that its post-training reduces generation from 50 denoising steps to six. In its technical deep dive, the company says its model led the evaluated systems in several categories, including human anatomy, complex composition and audio synchronization for text- and image-conditioned generation. The comparison used outputs at 768p and included MiniMax H3, Fal H3 Max and HeyGen Video; Pika also acknowledges weaker results in areas such as physics, multi-scene consistency and complex reference control.
The company says H3 Flex ranked first among open-weight models and seventh overall on DesignArena’s text-to-video leaderboard at launch. Those rankings reflect the launch snapshot, and Pika’s technical evaluation is company-reported rather than an independent assessment. The claims offer a benchmark context, but do not establish that every workflow or local hardware setup will produce the same results.
API access comes before the main creation app
Pika says H3 Flex is available through its API, with weights on Hugging Face and inference code on GitHub; the technical post also lists ComfyUI and Diffusers access. The model is available through Pika API Club, while the company says it plans to add Flex to its creation product on November 2. Developers can build on the released components under the model’s license terms, but Pika’s hosted finishing and serving features remain outside the local release.
For users deciding whether “open weight” means a complete self-hosted production pipeline, the boundary is concrete: local video generation with synchronized audio is included, but the advertised 4K finishing and five-stem audio export are not. H3 Flex’s release makes the generator available to run and modify, while leaving several of its most polished post-production steps on Pika’s service.