---

license: other license_name: minimax-h3-community license_link: LICENSE base_model:

• MiniMaxAI/MiniMax-H3 tags:
• android
• video
• quantized
• text-to-video
• image-to-video
• video-to-video
• tflite
• litert
• vae

---

Gestura-MiniMax-H3-VAE-LiteRT

Gestura app capabilities and measured performance

Generation mode	Input	256 × 256	384 × 384	Audio
Text-to-video (T2VA)	Text prompt	Preview, 5 s, 10 s	5 s	Yes
First-frame image-to-video (I2VA)	Text prompt + one image	Preview, 5 s, 10 s	5 s	Yes
Reference-conditioned video (REF2VA)	Text prompt + up to two reference inputs, at most one video	Preview, 5 s, 10 s (10 s: no video references)	—	Yes

App-level measurements on Samsung Galaxy S26 (SM-S9420), 256 × 256, 5-second output option (124 frames), 24 fps:

App input mode	Approximate generation time
Text → video with audio	~431 s (7 min 11 s)
First-frame image → video with audio	~460 s (7 min 40 s)
Two reference images → video with audio	~601 s (10 min 01 s)

Whole-app generation times; performance varies by input and device conditions.

Related models

Repository	Contents
FastH3 LiteRT	FastH3 denoising graphs and weights
Video VAE LiteRT (this repository)	INT8 video encoder and decoder
FastH3 GGUF	Q4_K_M denoising model
REF2VA Delta	Reference-conditioned generation weight-delta package
Video VAE GGUF	Q8_0 video VAE

On-device video examples

Four screen recordings from Gestura on Samsung Galaxy S26, showing generated video with audio.

Download Gestura

Gestura is an Android app for on-device AI generation, including text-to-video, first-frame image-to-video and reference-conditioned video with audio. The models in these repositories were converted or packaged during its development.

Get Gestura on Google Play

Model overview

INT8 video VAE encoder and decoder graphs converted from MiniMax H3. The encoder maps reference frames into video latents; the decoder turns generated latents into video frames. The split graphs allow applications to execute the VAE in stages.

These assets were converted or packaged during Gestura development and are available for integration into other projects under the licenses below.

Property	Value
Format	TensorFlow Lite / INT8 weights
Model payload size	3.01 GB (decimal)
Model payload files	82
Integrity metadata	files.json, with sizes and SHA-256 hashes

Repository contents

Path	Contents
int8-source-direct-v1/	Split decoder graphs
int8-causal-encoder-v1/	Causal encoder graphs with 1-, 5- and 17-frame variants
int8-full384-shared-v2/	Decoder graph set for 384-resolution processing

Model usage

Load the encoder or decoder graph set with LiteRT CompiledModel. Inspect each graph signature for tensor names, shapes and dtypes, and connect the split stages in graph order. The encoder uses THWC layout and produces posterior-mean latents. Keep any external weight files beside their graph set. A host pipeline supplies latent normalization, temporal state and frame assembly; this repository contains the VAE assets.

License and attribution

The model assets remain subject to the MiniMax H3 Community License Agreement, including its Acceptable Use Policy (Exhibit A). Read the complete LICENSE and NOTICE; this summary does not replace them.

• Territory restrictions exclude the European Union, United Kingdom, Republic of Korea and United States; separate authorization is needed for uses not granted by the agreement.
• Annual commercial revenue above USD 20 million requires prior written MiniMax authorization. Commercial products/services must prominently display MiniMax H3 in their UI.
• Redistribution requires the agreement, attribution and modification notices. Downstream terms, access/use safeguards and abuse reporting obligations also apply.
• The agreement restricts improving other AI models, imposes acceptable-use rules, and requires disclosure of publicly shared AI-generated content. Warranty, liability, termination and other provisions remain applicable.

Converted or packaged by Gestura; original model rights remain with the respective upstream owners. No upstream endorsement is implied.

Upstream: MiniMaxAI/MiniMax-H3 · ChrisColeTech/minimax-h3-turbo-GGUF.