You need to agree to share your contact information to access this model
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
Access is available only in the licensed Territory. By requesting access, you agree to the Tencent Hunyuan Community License Agreement, its Acceptable Use Policy, and all restrictions described below.
Log in or Sign Up to review the conditions and access this model content.
PocketAI Video MLX — HunyuanVideo 1.5
PocketAI Video MLX is an Apple-Silicon-native MLX edition of Tencent HunyuanVideo 1.5's distilled 480p text-to-video model. It combines PocketAI's MLX implementation, conversion, mixed-precision quantization, packaging, and runtime with the original HunyuanVideo 1.5 model design and weights.
Powered by Tencent Hunyuan. This is a community derivative and is not an official Tencent release.
Precision editions
All PocketAI MLX editions are published in this single gated repository. Pick one revision so only that edition is downloaded:
| Edition | Revision | Package size | Recommended use |
|---|---|---|---|
| Q4 | v1.1.0-q4 |
14.31 GB | Lowest download and weight footprint |
| Q8 | v1.1.0-q8 |
21.08 GB | Recommended balance of fidelity and size |
| BF16 | v1.1.0-bf16 |
31.62 GB | Full-fidelity MLX reference |
The default main revision contains Q8. The moving q4, q8, and bf16
branches are also available, while the versioned tags above are recommended
for reproducible installs.
Important license and territory restriction
This release is governed by the Tencent Hunyuan Community License Agreement. The licensed Territory excludes the European Union, United Kingdom, and South Korea. Do not download, access, use, distribute, display, or make model outputs available outside the licensed Territory. The full license and Acceptable Use Policy control over this summary.
This is a source-available/open-weight community release, not an OSI-approved open-source model.
What is included
| Component | Format | Size | Notes |
|---|---|---|---|
transformer/ |
MLX 8-bit group-32 + sensitive BF16 paths | 11.90 GB | 480p T2V distilled DiT |
text_encoder/ |
MLX 8-bit group-32 | 7.44 GB | Required 26-layer Qwen2.5-VL text tower and tokenizer |
vae/ |
MLX BF16 | 1.74 GB | Decoder-only causal 3D VAE |
src/pocketai_video/ |
Python/MLX | — | Standalone generation runtime |
The optional Glyph-SDXL-v2 ByT5 checkpoint is deliberately not included because its upstream redistribution license is not declared. Normal prompts work. Prompts containing quoted display text require that optional encoder.
Requirements
- Apple Silicon Mac
- Python 3.11
ffmpegavailable onPATH- 24 GB or more unified memory recommended
Install
First request and receive access to this gated repository. Then:
hf auth login
hf download PocketAiHub/PocketAI-Video-MLX-HunyuanVideo-1.5 \
--revision v1.1.0-q8 \
--local-dir pocketai-video-mlx
cd pocketai-video-mlx
python3.11 -m venv .venv
.venv/bin/pip install .
Install ffmpeg if needed:
brew install ffmpeg
Generate a video
The default is a practical 832×480, 9-frame, 10-step first render:
.venv/bin/pocketai-video "A red fox runs through a snowy forest, cinematic side-tracking shot." --output fox.mp4
For a quick lower-resolution check:
.venv/bin/pocketai-video "Ocean waves roll beneath a moonlit sky." --height 256 --width 448 --frames 9 --steps 10 --output waves.mp4
Validate a downloaded snapshot without loading the weights:
.venv/bin/pocketai-video "validation" --validate-only
The command writes both the MP4 and an adjacent JSON generation report.
Performance and fidelity
On the development Apple Silicon Mac, the quality tier completed an 832×480, 9-frame, 10-step render in 62.9 seconds including denoising and VAE decode. At 128×224×9 with classifier-free guidance and 10 steps, the quality-tier denoiser took 2.28 seconds, used 12.72 GiB active memory, and peaked at 13.47 GiB.
The quality transformer has 0.99599 single-prediction cosine similarity to the BF16 reference. The quality Qwen tower has 0.999917 cosine similarity on active real-prompt embeddings.
Scope and limitations
- Text-to-video only in this first release.
- The released VAE contains the decoder only.
- The Qwen export excludes the unused vision tower, LM head, final norm, and two unused language layers.
- Quoted display-text prompts require the separately licensed Glyph ByT5 encoder and are rejected when it is absent.
- Quantized diffusion runs are deterministic for a given environment and seed, but are not expected to be pixel-identical to BF16.
Provenance
- Tencent source:
Tencent-Hunyuan/HunyuanVideo-1.5at60783e704160023913bee78f0b47036d393d4dfa - Tencent checkpoint:
tencent/HunyuanVideo-1.5at9b49404b3f5df2a8f0b31df27a0c7ab872e7b038 - Qwen text tower:
Qwen/Qwen2.5-VL-7B-Instructatcc594898137f460bfe9f0759e9844b3ce807cfb5 - Complete source hashes and component details:
pocketai-release-manifest.json - PocketAI changes:
MODIFICATIONS.md - Third-party notices:
THIRD_PARTY_NOTICES.md
Attribution
Tencent Hunyuan is licensed under the Tencent Hunyuan Community License Agreement, Copyright © 2025 Tencent. All Rights Reserved. The trademark rights of “Tencent Hunyuan” are owned by Tencent or its affiliate.
Qwen2.5-VL-derived text-encoder files remain under Apache-2.0. See
LICENSE-QWEN.
Quantized
Model tree for PocketAiHub/PocketAI-Video-MLX-HunyuanVideo-1.5
Base model
tencent/HunyuanVideo-1.5