Access PocketAI Video MLX — LTX-2.3

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

This repository contains a derivative of LTX-2. By accessing it, you agree to the complete LTX-2 Community License Agreement and Attachment A included in this repository. The full agreement controls over this summary.

PocketAI Video MLX — LTX-2.3 Distilled Edition

PocketAI Video MLX is an Apple-Silicon-native, 4-bit MLX edition of Lightricks' LTX-2.3 22B distilled 1.1 audio-video model. It generates video and synchronized audio locally on a Mac and includes the two-stage spatial upscaler path.

Powered by LTX-2 from Lightricks. This is a community derivative released by PocketAI and is not an official Lightricks release. No endorsement is implied.

License

This release is governed by the complete LTX-2 Community License Agreement, including Attachment A. The agreement permits derivatives and redistribution subject to its terms. Among other requirements, entities with aggregated annual revenue of at least USD 10,000,000 must obtain the paid commercial-use license required by Lightricks. Export-control, sanctions, use, disclosure, and downstream distribution requirements also apply.

Read the complete agreement before using or redistributing this release. This is a source-available/open-weight community model, not an OSI-approved open-source model.

What is included

Component	Format	Approximate size	Notes
Distilled 1.1 transformer	MLX 4-bit, group size 64	11.32 GB	48-layer LTX-2.3 audio-video DiT
Embeddings connector	MLX safetensors	6.34 GB	Video and audio prompt projections
Video VAE encoder and decoder	MLX safetensors	1.45 GB	Image/video conditioning and decode
Spatial upscaler	MLX safetensors	1.00 GB	Distilled two-stage ×2 path
Audio VAE and vocoder	MLX safetensors	365 MB	Synchronized audio generation
Native runtime	Python and MLX	less than 2 MB	Pinned `ltx-2-mlx` runtime source

Model files total approximately 20.48 GB.

The Gemma 3 12B 4-bit prompt encoder is not duplicated in this repository. The runtime downloads mlx-community/gemma-3-12b-it-4bit on first use. You must separately accept Google's Gemma license and will need approximately 7.5 GB of additional disk space.

Requirements

Apple Silicon Mac
Python 3.11, 3.12, or 3.13
ffmpeg available on PATH
16 GB unified memory with --low-ram; 24 GB or more recommended
Approximately 30 GB free disk space including the separate Gemma encoder

Install

Request access to this repository and accept the Gemma terms first. Then:

hf auth login
hf download PocketAiHub/PocketAI-Video-MLX-LTX-2.3 --local-dir pocketai-ltx-2.3-mlx
cd pocketai-ltx-2.3-mlx
python3 -m venv .venv
.venv/bin/pip install .

Install ffmpeg if needed:

brew install ffmpeg

Generate a video with synchronized audio

.venv/bin/pocketai-ltx-video generate \
  --model . \
  --gemma mlx-community/gemma-3-12b-it-4bit \
  --distilled \
  --prompt "A red fox runs through a snowy forest, cinematic tracking shot." \
  --height 480 --width 704 --frames 97 --frame-rate 24 \
  --stage1-steps 8 --stage2-steps 3 \
  --seed 42 --low-ram \
  --output fox.mp4

For a quick functional check:

.venv/bin/pocketai-ltx-video generate \
  --model . \
  --gemma mlx-community/gemma-3-12b-it-4bit \
  --distilled \
  --prompt "A paper boat glides across calm water." \
  --height 256 --width 256 --frames 17 --frame-rate 24 \
  --stage1-steps 1 --stage2-steps 1 \
  --seed 2301 --low-ram \
  --output smoke.mp4

LTX's temporal VAE works best with frame counts of 8N+1 such as 17, 49, or 97. Width and height should be divisible by 32.

Release validation

PocketAI validated the exact packaged weights with MLX Forge and then ran the bundled native pipeline end to end:

256×256, 17 frames, 24 fps
Distilled two-stage path, 8 first-stage and 3 refinement steps
4-bit transformer with low-RAM streaming
Gemma prompt encoding, video VAE, audio VAE, vocoder, spatial upscaler
Valid H.264 video and AAC audio output
24.0 seconds total on the development Apple Silicon Mac

Performance varies by Mac, memory pressure, resolution, and frame count.

Scope and limitations

This checkpoint supports the --distilled pipeline. It does not contain the separate development transformer required by the one-stage and CFG modes.
The separate Gemma prompt encoder is required and carries its own license.
The first generation downloads Gemma and is slower than subsequent runs.
Quantization trades some numerical fidelity for substantially lower memory use and faster local execution.
Video generation can produce visual, temporal, prompt-following, and audio artifacts. Review outputs before publication.
Generated content must be identified as machine generated where required by Attachment A of the LTX-2 license.

Provenance

Base model: Lightricks/LTX-2.3 at 4229404625088d21c4f112eb640fb04a0900ee25
Base checkpoint: ltx-2.3-22b-distilled-1.1.safetensors
Spatial upscaler: ltx-2.3-spatial-upscaler-x2-1.1.safetensors
Conversion tool: mlx-forge v0.3.4 at 53a2de9f059efd1b1c527a383d51b4ff8c464ff2
Runtime: dgrauet/ltx-2-mlx v0.14.18 source at 36ab612919ccba4b90321019a2927633dd837d35
Quantization: affine 4-bit, group size 64, transformer block linear layers
Complete hashes: pocketai-release-manifest.json
Changes: MODIFICATIONS.md
Notices: THIRD_PARTY_NOTICES.md

Downloads last month: 6

MLX

Hardware compatibility

Quantized

Model tree for PocketAiHub/PocketAI-Video-MLX-LTX-2.3

Base model

Lightricks/LTX-2.3

Quantized

(59)

this model