AI video startup Pika has released H3 Flex, an open-weight video generation model designed for professional creators, developers and researchers. Announced on October 29, 2026, the model generates video with synchronised audio, supports up to 12 reference inputs and uses a six-step generation process designed to reduce production time.
Built by post-training MiniMax H3, Pika H3 Flex supports text-to-video, image-to-video and multimodal reference-based generation. It produces native video at 480p and 768p, with Pika’s hosted services extending output to 1080p, 2K and 4K. The company has also released model weights and inference code, allowing developers to experiment with the technology and integrate it into their own workflows.
Key Takeaways
- Pika launched H3 Flex on October 29, 2026.
- The model is a post-trained version of MiniMax H3.
- It uses six generation steps instead of the original 50-step process.
- It supports up to 12 reference files across images, videos and audio.
- Native video generation supports 480p and 768p resolutions.
- Pika’s hosted API offers finishing options up to 4K.
- Generated videos include synchronised stereo audio.
- Five separate audio stems are available through Pika’s hosted API.
- Model adapters are available on Hugging Face, with inference code on GitHub.
What Is Pika H3 Flex?
Pika H3 Flex is an open-weight AI model designed to generate and edit videos using text prompts, images, videos and audio references.
The model builds on MiniMax H3, which Pika has post-trained to improve generation speed, visual quality and creative control.
Unlike conventional text-to-video systems that primarily generate footage from written prompts, H3 Flex supports multimodal references. Creators can provide images of characters, product photographs, sketches, moodboards and video clips to guide the generated result.
For example, a filmmaker could provide:
- A character reference image.
- A photograph of the desired environment.
- A video showing the intended camera movement.
- An audio reference for the desired sound.
The model can use these inputs together to generate a video that follows the supplied creative direction.
Pika is positioning H3 Flex as a tool for professional iteration rather than simply generating one video from one prompt.
Six-Step Generation Makes Video Creation Faster
One of the most notable technical changes is the model’s six-step generation process.
Pika says it post-trained MiniMax H3 to generate videos using six steps instead of 50. The company then refined the model to improve visual quality and creative control.
According to Pika, H3 Flex can generate a 15-second video at 768p with synchronised audio in approximately seven seconds of transformer computation under a controlled benchmark.
However, this figure measures the transformer’s computation time, not the complete end-to-end generation time. Reference processing, decoding, service overhead and higher-resolution finishing can increase the actual waiting time.
The reduced sampling process is intended to make experimentation faster. Creators can test different prompts, camera movements, characters and visual styles with less time between iterations.
H3 Flex Supports Up to 12 Reference Inputs
Another major feature is its multimodal reference system.
H3 Flex accepts up to 12 reference files, combining images, videos and standalone audio inputs within supported limits.
| Input type | Supported references |
|---|---|
| Images | Up to 9 |
| Videos | Up to 3 |
| Standalone audio clips | Up to 3 |
| Total reference files | Up to 12 |
The model allows creators to use different references for different elements of a scene.
For example, one image can define a character, another can establish the environment, while a video can guide choreography or camera movement.
The system also supports first-frame and last-frame conditioning, allowing creators to provide opening and ending images to guide the generated sequence.
This makes H3 Flex useful for product videos, animation, advertisements, motion graphics and multi-shot storytelling.
Video Generation Supports Up to 4K Resolution
Pika H3 Flex generates video natively at 480p and 768p.
For higher-resolution output, Pika provides a separate hosted upsampling service that can extend generated footage to 1080p, 2K and 4K.
| Feature | Specification |
|---|---|
| Native resolutions | 480p and 768p |
| Higher-resolution finishing | 1080p, 2K and 4K |
| Video duration | 5–15 seconds |
| Frame rate | 24 fps |
| Aspect ratios | 11 supported formats |
| Audio | 32 kHz stereo |
| Output format | MP4 with H.264 video and AAC audio |
The distinction between native generation and upscaled output is important. The downloadable model does not independently generate native 4K video. Higher-resolution finishing is offered through Pika’s hosted services.
Creators can therefore generate lower-resolution drafts, select their preferred version and apply the finishing process to the final clip.
Pika Introduces Five Separate Audio Stems
H3 Flex generates video with synchronised stereo audio.
Pika also offers a hosted audio-separation feature that divides the generated soundtrack into five editable components:
- Dialogue: Spoken words and conversations.
- Score: Background music.
- Sound effects: Designed effects associated with actions and events.
- Foley: Everyday sounds such as footsteps and object movements.
- Ambience: Background environmental audio.
This feature can help video editors adjust individual sound elements without having to recreate the entire soundtrack.
For example, an editor could lower the background music while keeping dialogue clear or modify environmental sounds independently.
The five-stem separation feature is provided through Pika’s hosted API rather than the downloadable local inference package.
Pika Releases Model Weights on Hugging Face
Pika is making H3 Flex available beyond its own creation platform.
The company has released its model adapters on <Entity value=”Hugging Face” category=”software” disambig=”AI model hosting and development platform”/> and inference code through <Entity value=”GitHub” category=”software” disambig=”Code hosting and collaboration platform”/>.
Developers can use the released components to run inference, experiment with model behaviour and integrate video generation into their own applications.
The model supports integration through tools including ComfyUI and Diffusers.
However, open-weight availability does not mean every feature is available for local deployment. The local package supports native 480p and 768p generation with mixed stereo audio, while higher-resolution finishing, audio-stem separation and some additional features remain Pika-hosted services.
The model adapters are distributed under the MiniMax H3 Community License Agreement. The inference and loader code uses Apache 2.0 licensing. Developers should review the relevant model and component licenses before commercial deployment.
How H3 Flex Compares With Traditional AI Video Workflows
Traditional video production often requires multiple tools for generating footage, editing visuals, creating sound effects and preparing final output.
H3 Flex attempts to bring several of these steps closer together.
| Capability | H3 Flex |
|---|---|
| Text-to-video | Supported |
| Image-to-video | Supported |
| First/last-frame conditioning | Supported |
| Video reference guidance | Supported |
| Audio reference guidance | Supported |
| Synchronized audio generation | Supported |
| Multi-shot storytelling | Supported |
| Motion transfer | Supported |
| Object replacement | Supported |
| Character replacement | Supported |
| 4K finishing | Available through hosted service |
| Separate audio stems | Available through hosted service |
Pika’s technical evaluation reports improvements against selected competing models across several visual and audio metrics. These results come from Pika’s own evaluation and should not be interpreted as an independent industry-wide ranking.
The company also acknowledges that areas such as physics, multi-scene consistency and complex reference control still require improvement.
The Bigger Picture
The release of H3 Flex highlights an important development in the AI video market: companies are increasingly competing on creative control, speed and accessibility rather than video generation alone.
Open-weight models give developers more opportunities to customise generation systems and build specialised applications. Multimodal references also address a major limitation of prompt-only video generation by allowing creators to provide visual and motion guidance directly.
Pika’s approach combines downloadable model components with hosted services for higher-resolution output and advanced audio editing. This hybrid model gives developers access to the underlying generation technology while allowing Pika to offer additional capabilities through its infrastructure.
Looking Ahead
The next challenge for H3 Flex will be adoption among developers and professional creators. Its open-weight release could encourage experimentation in animation, advertising, product videos and AI-powered editing applications. Hardware requirements, licensing conditions and the quality of locally generated output will influence how widely developers adopt it.
Competition in AI video generation is also likely to intensify as models become faster and more controllable. H3 Flex gives Pika a stronger presence in the open-weight segment, but long-term adoption will depend on consistent output quality, reliable reference handling and the ability to maintain visual continuity across complex scenes.
Frequently Asked Questions
1. What is Pika H3 Flex?
Pika H3 Flex is an open-weight AI video and audio generation model developed by post-training MiniMax H3. It supports text-to-video, image-to-video and multimodal reference-based generation.
2. When was Pika H3 Flex released?
Pika announced H3 Flex on October 29, 2026.
3. Can Pika H3 Flex generate 4K videos?
The model generates natively at 480p and 768p. Pika offers hosted upsampling services to produce 1080p, 2K and 4K output.
4. How many reference images and videos does H3 Flex support?
Its reference system supports up to 12 files in total, including up to nine images, three videos and three standalone audio clips, subject to the combined reference limits.
5. Is Pika H3 Flex open source?
Pika has released open weights and inference code. The model adapters are distributed under the MiniMax H3 Community License Agreement, while the inference code uses Apache 2.0. It is therefore important to distinguish open-weight availability from unrestricted open-source licensing.
6. Where can developers access Pika H3 Flex?
Developers can access the model weights through Hugging Face and the inference code through GitHub. It is also available through Pika’s API and supports ComfyUI and Diffusers integrations.
7. Does H3 Flex generate audio?
Yes. It generates synchronised stereo audio alongside video. Pika’s hosted API can additionally provide five separate audio stems: dialogue, score, sound effects, foley and ambience.
Get the day’s top stories in your inbox
One concise email. No spam, unsubscribe anytime.
