Karam98/ltx-2.5-mlx-q4

LTX-2.5 · MLX · 4-bit ( mirror)

The LTX-2.5 render pack that nstalls for Apple Silicon: the diffusion transformer quantised to 4-bit (group size 64, AdaLN / patch embedding / output projections kept bf16), plus the connector, VAE encoder/decoder, audio VAE, vocoder, the x2 spatial and temporal upscalers and the duration head, in the flat single-file-per-component layout the ltx-2-mlx runtime loads.

This is an unmodified redistribution of Phosphene's ltx-2.5-mlx-q4 pack. Every byte, including the provenance metadata inside each safetensors file, is the pack as Phosphene built it; phosphene_quant_manifest.json carries the sha256 of every file and was verified before upload.

Attribution

Original model	LTX-2.5 — Lightricks Ltd. (https://huggingface.co/Lightricks/LTX-2.5)
MLX runtime	dgrauet/ltx-2-mlx (MIT)
Conversion + quantisation	Phosphene — https://github.com/mrbizarro/phosphene — see NOTICE.md for the full statement of changes
This mirror	— a byte-identical copy, so the pack stays downloadable for installs

Licence — read before use

These weights are a Derivative of LTX-2.x and are distributed exclusively under the LTX-2.x Community License Agreement, a complete copy of which is in this repo as LICENSE-LTX-2.x-Community-License.md. By downloading or using them you accept that Agreement in full, including the use restrictions in Section 4 and Attachment A, and you must pass those terms on to anyone you distribute these weights — or any derivative of them — to. Entities with annual revenues of US $10,000,000 or more need a paid licence from Lightricks for any non-personal use. NOTICE.md (from Phosphene) spells out the changes made to the upstream release and the terms passed on.