Gemini Omni Flash
Model Card

 ​
 ​
Model Cards are intended to provide developers with essential, summarized information on
models, including overviews of known limitations and mitigation approaches. Model cards may be
updated from time to time; for example, to include updated evaluations as the model is improved
or revised. See the Google DeepMind site for a comprehensive list of model cards.

Initially Published: May 2026,
Last Updated: August 2026

                                         Model Information

Description: Gemini Omni Flash is our next step towards models that can create and edit anything from
any input—starting with video. It combines Gemini’s intelligence with our generative media models,
representing a leap forward in world understanding, multimodality, and editing. Gemini Omni Flash
enables high-quality video creation and a more natural way to edit videos through conversation. This
model card describes Gemini Omni Flash and Gemini Omni 1.1 Flash.

Inputs: Text strings (e.g., a question, a prompt), images, audio, and video files.

Outputs: High-quality, high-resolution video with audio.

Architecture: Gemini Omni Flash is a transformer-based model (Vaswani et al., 2017) with native
multimodal support for text, vision, video and audio inputs.

                                              Model Data

Training Dataset: Gemini Omni Flash was trained on audio, video, image, and text data. Audio and video
datasets were annotated with text captions at different levels of detail.

Training Data Processing: Training videos were also filtered for various compliance, safety, and quality
metrics and deduplicated semantically.

                             Implementation and Sustainability

Hardware: Gemini Omni Flash was trained using Google’s Tensor Processing Units (TPUs). TPUs are
specifically designed to handle the massive computations involved in training LLMs and can speed up
training considerably compared to CPUs. TPUs often come with large amounts of high-bandwidth
memory, allowing for the handling of large models and batch sizes during training, which can lead to

                                                                                                           1
better model quality. TPU Pods (large clusters of TPUs) also provide a scalable solution for handling the
growing complexity of large foundation models. Training can be distributed across multiple TPU devices
for faster and more efficient processing.

The efficiencies gained through the use of TPUs are aligned with Google's commitment to operate
sustainably.

Software: Training was done using JAX and ML Pathways.

                                              Distribution
Gemini Omni Flash is distributed in the following channels; respective documentation shared in line:

    ●​   Gemini App
    ●​   YouTube
    ●​   Google Flow
    ●​   Google Flow Music

                               Intended Usage and Limitations

Benefit and Intended Usage: Gemini Omni Flash can be used to generate high-quality, high-resolution
videos from any input in a wide range of visual styles. The model is able to faithfully follow simple and
complex instructions, simulate real-world physics, and edit videos through conversation.

Known Limitations: While Gemini Omni Flash demonstrates strong progress, maintaining complete
consistency throughout edits, generating scenes with complex motion, or rendering perfectly accurate
text remains a challenge.

Acceptable Usage: Google’s Generative AI Prohibited Use Policy applies to uses of the model in
accordance with the applicable terms of service. Additionally, the model should not be integrated into
certain systems (also found in Google’s Generative AI Prohibited Use Policy), including those that: (1)
engage in dangerous or illicit activities, or otherwise violate applicable laws or regulations, (2)
compromise the security of others’ or Google’s services, (3) engage in sexually explicit, violent, hateful,
or harmful activities, (4) engage in misinformation, misrepresentation, or misleading activities.

                                                                                                              2
                                    Ethics and Content Safety

Evaluation Approach: Gemini Omni Flash was developed in partnership with internal safety, security,
and responsibility teams. A range of evaluations and red teaming activities were conducted to help
improve the model safety and inform decision-making. These evaluations and activities align with
Google's AI Principles and responsible AI approach, as well as Google's Generative AI policies (e.g. Gen
AI Prohibited Use Policy and the Gemini API Additional Terms of Service). Evaluation types included but
were not limited to:

    ●​ Training/Development Evaluations including automated and human evaluations carried out
       continuously throughout and after the model’s training, to monitor its progress and
       performance;
    ●​ Human Red Teaming conducted by specialist teams who sit outside of the model development
       team, across the policies and desiderata, deliberately trying to spot weaknesses and ensure the
       model adheres to safety policies and desired outcomes;
    ●​ Automated Red Teaming to dynamically evaluate Gemini Omni Flash for safety and security
       considerations at scale, complementing human red teaming and static evaluations;
    ●​ Ethics & Safety Reviews conducted ahead of the model’s release.

Responsible Innovation: AI multimodal generation and editing tools can help lower barriers to
entry and transform education through personalized audio-visual content. Beyond direct
applications, its impact extends to advanced research.

By producing high-quality synthetic data, video generation can help accelerate innovation in
robotics, computer vision, and generative 3D technologies.

Such advanced multimodal creation capabilities require a proactive approach to safety.

We focused on two main content safety areas:​
​       (i) Intentional adversarial misuse of the model; and,
        (ii) Unintentional model failure modes through benign use.

Similar to our safety approach to our other Gemini models, we built safety at the core of Gemini Flash
Omni:
    ●​ Pre-training mitigations: We used diverse synthetic captioning to improve the model’s ability to
       accurately represent a wide variety of concepts
    ●​ Post-training mitigations: To maintain high information integrity and prevent unwanted
       outputs, we deployed advanced production filters and used SynthID, our digital watermarking
       tool to clearly verify AI-generated content

As part of editing videos, Gemini Omni Flash is capable of changing people's speech. For
now, we are restricting this capability and working to better understand how to safely and
responsibly bring it to our users

                                                                                                           3