You need to agree to share your contact information to access this model
By clicking "Agree and Access," you agree to the LTX-2 Community License Agreement and Acceptable Use Policy.
Log in or Sign Up to review the conditions and access this model content.
LTX-2.5 Distilled MLX — BF16
Community Apple-Silicon-native MLX conversion of the official Lightricks LTX-2.5 distilled audio-video pipeline, produced and validated by PocketAI Model Lab. PocketAI did not train or fine-tune the underlying model. This is not an official Lightricks release, and no endorsement is implied.
The conversion implements the complete 8-step ancestral first stage and 3-step spatial refinement stage, Gemma 4 text conditioning, the convolutional video VAE, and synchronized stereo audio.
This candidate contains 70,990,089,274 bytes of model payload: 47,143,115,060
bytes in the main snapshot and 23,846,974,214 bytes beneath
gemma4-text-encoder/.
Validation
On an Apple M5 Max with 128 GB unified memory, the matched BF16 smoke took 23.24 seconds and peaked at 39.75 GB of Metal allocation at 384×256 and 17 frames. Q8 also passed the official 768×512, 121-frame, 24 fps preset in 93.14 seconds at a 22.89 GB peak, with H.264 video and synchronized 48 kHz stereo AAC audio.
Q8 is the recommended catalog candidate. Q4 is a compact, provisional option with greater composition drift. BF16 is the high-memory reference.
Runtime
The release is pinned to dgrauet/ltx-2-mlx revision
e1838a855bfd1640135c424c96cb27a0c0ad150e plus the seven patches in
runtime-patches/. PocketAI desktop integration must consume that exact runtime
contract; the current LTX-2.3 runtime is not sufficient for this artifact.
License and access terms
The complete upstream agreement is included as
LICENSE-LTX-2.5.md and is also available from the
pinned official Lightricks source.
The current official Acceptable Use Policy,
last updated March 30, 2026, is bundled as
ACCEPTABLE-USE-POLICY.pdf.
MODIFICATIONS.md records PocketAI's derivative changes.
The model page is public for transparency and discovery. Model-file access is granted automatically after the user clicks Agree and Access; no identity, organization, revenue, or intended-use questionnaire is collected. Access does not alter or replace the upstream agreement.
- Downloads last month
- 12
Quantized
Model tree for PocketAiHub/LTX-2.5-MLX
Base model
Lightricks/LTX-2.5