Latest in AI
HexAnima and MiniMax H3 Image: Check Your Phone Before Downloading 48 GB
The new Android model pack makes local generation easier to try, but device support and licensing rule out many seemingly suitable setups.
HexAnima 1.6.0 adds a downloadable MiniMax H3 Image pack for local Android generation. Released on November 14, the independent project makes a large image model easier to try on a phone, but the 48 GB download is only worth considering after checking hardware support and the model’s licence.
This is a requirements guide based on the project’s documentation, not a phone test. The important question is whether your exact setup is supported, not whether a model with 33 billion parameters can technically run somewhere on Android.
Which Android phones does HexAnima support?
The project’s requirements name Snapdragon 8 Gen 1 and 8+ Gen 1, using Hexagon V69. Android 10 or newer and a 64-bit system are required. No root is needed.
Newer does not automatically mean compatible. The same documentation says Snapdragon 8 Gen 2, 8 Gen 3 and 8 Elite are unsupported. Experimental 8 Elite support was withdrawn after problems on a tester’s Galaxy S25. The documented test phone has 12 GB RAM; general low-memory support elsewhere in the app is not proof that H3 works on an 8 GB device.
Check the chipset rather than relying on the phone’s marketing name. A newer flagship can be the wrong device for a project built around a particular accelerator. Buying a replacement phone for this release would be premature without a confirmed result for that exact model and software version.
Check the model licence before the download
HexAnima’s code licence and the downloaded model’s licence are different things. The licence shipped with the H3 pack limits its grant to a territory excluding the United States, European Union, United Kingdom and Republic of Korea. It directs interested users in excluded territories to contact MiniMax about separate licensing.
That is a reason to stop and resolve permission, not to look for another download mirror. The agreement also prohibits using the model or its outputs to improve other AI models, except H3 Image itself or its model derivatives. These are selected terms, not a complete legal assessment or a determination that a particular use is permitted. Read the full agreement for your intended workflow.
An open repository does not make every bundled model unrestricted. Keep the model licence with your evaluation notes instead of assuming the app’s availability settles the question.
What changed in version 1.6.0?
The versioned release adds H3 to catalog8.json; older app versions continue reading catalog7.json. Its pack is hosted separately as ayanoby/hexanima-h3, with a manifest containing file sizes and SHA-256 hashes. The release describes INT8-ConvRot weights, six-step generation without CFG and an optional two-step pDMD LoRA.
If H3 is missing from an existing installation, check the app version before assuming the model has disappeared. Equally, a downloaded file is not a verified working installation. File-integrity checks and successful generation answer different questions.
The practical gain is easier access to a packaged conversion, rather than a new MiniMax model release. It does not establish that this conversion matches the original weights’ quality or that the fastest settings preserve every detail.
Plan storage and evaluate one small job first
The project reports streaming H3’s weights from storage for each picture and roughly two minutes for a 512-by-768 result. That is the maintainer’s reported performance, not a result reproduced for this article. A 48 GB download describes storage, not a 48 GB RAM requirement.
For a first evaluation, choose one ordinary prompt and a modest output size, then record the app version, selected model, settings, elapsed time and whether the result meets the brief. Keep a failed run in the record. Do not compare a carefully selected image from one setup with the first output from another and call that a quality test.
If you only have enough free space for the pack itself, resolve the storage budget before starting. The download is not your entire working set: the application, generated images and other installed models also need room. There is no verified universal free-space allowance or H3 memory minimum established here.
Local generation is useful when it fits the device you already own and the work you are allowed to do. For this release, compatibility and permission are the decision; parameter count is background information.
I’m Paras. I work at the intersection of AI, engineering and distribution.
I help technical teams turn model releases and engineering trade-offs into clear, useful developer stories.