ltx-2.3-mlx-q4

ltx-2.3-mlx-q4 — int4 distilled transformer

A verbatim availability mirror. The distilled transformer quantized to 4-bit (group 64, transformer blocks only). Pairs with the components from xocialize/ltx-2.3-mlx.

Why this repo exists

This is an unmodified mirror of dgrauet/ltx-2.3-mlx-q4, hosted so that downstream software has a weight source that does not depend on a third-party namespace remaining available. It exists for durability, not improvement — there is nothing here you cannot get upstream, and upstream is the better place to get it if you want the newest revision.

The motivation is concrete rather than hypothetical: an unrelated model family we ship ports of was relocated to a new organization without HTTP redirects, and every published package referencing the old paths broke the same day. Availability is a dependency like any other.

Provenance chain

We did not convert, quantize, fine-tune, or otherwise alter these weights. Every tensor file here was verified by SHA-256 against the corresponding upstream file at mirror time; the hashes are listed below so anyone can check that claim independently.

Small JSON metadata files were taken fresh from upstream at mirror time and may therefore carry metadata additions relative to an older local copy; the tensor files are the verified artifacts.

Licence — read this before use

These weights are governed by the LTX-2 Community License Agreement (dated January 5, 2026), included in full as LICENSE in this repository. Mirroring does not alter, relax, or re-license anything.

Two obligations deserve to be called out explicitly, because they are easy to miss and they bind you as a user of these weights, not just us as a distributor:

1. Use-based restrictions (paragraph 4 and Attachment A) apply in their entirety. Attachment A is an enforceable part of the licence and lists prohibited uses. You must comply with it, and you must require the same of anyone you distribute to.
2. Attachment A item 20 is a non-compete. Using LTX-2 or its derivatives in a product or service that competes with, replaces, or substitutes Lightricks' commercial offerings requires a separate commercial licence from Lightricks. There is also a revenue threshold in the agreement's commercial terms. If you are building a product, read the licence yourself rather than relying on this summary.

Any redistribution of these weights or derivatives of them must carry a complete copy of the agreement and propagate these terms (§3). This summary is a courtesy pointer, not a substitute for the licence text, and where the two differ the licence governs.

Verified file hashes (SHA-256)

Each matches the corresponding file at dgrauet/ltx-2.3-mlx-q4 as of 2026-08-03:

8e17eec7f042aafde2c1d8824e16fa01d9c4be62bf1d36730e7e123080284196  transformer-distilled.safetensors

Credit

All model design, training, and weights are the work of Lightricks. The MLX conversion and the accompanying inference pipeline are the work of dgrauet. This mirror contributes nothing but disk space and an address that we can promise not to move.