LTX-2.5 β€” int8 DiT (MLX)

The transformer only, quantized to int8 group-64, for mlx-community/ltx-2.5-mlx. Every other component (Gemma-4 text encoder, connector, VAEs, vocoder, duration head, upscalers) comes from that base repo β€” this repo deliberately carries one file, so there is nothing to keep in sync.

MLX conversion of Lightricks LTX-2.5, consumed by ltx-2-mlx (Python) and ltx-2-mlx-swift (Swift).

Why this exists

LTX-2.5's bf16 DiT is ~38 GB resident. On a 64 GB machine that left the tier it ships on at 96% of its memory budget β€” technically admissible, with 1.69 GB of headroom, at a geometry below the profile's own frame cap. int8 takes the same tier to 57%.

Recipe

int8, group 64, transformer_blocks only β€” identical to the LTX-2.3 recipe, and read off the shipped 2.3 q8 checkpoint rather than invented. QUANT-RECIPE.json in this repo records it machine-readably.

⚠️ There is no int4 sibling, deliberately. int4 was built, gated and rejected: connector output at valid token positions read 0.996728 against a 0.999879 bf16 floor β€” 27Γ— the baseline angular error. It is not published because it should not be used.

Fidelity

Measured against the bf16 DiT on identical inputs at the shipping geometry (N=5632): cosine 0.99824 β€” and for calibration, LTX-2.3's q8 measured the same way reads 0.99238, so 2.5's int8 sits 4.3Γ— closer to its bf16 than 2.3's does.

⚠️ A quant-vs-bf16 cosine is only readable against the same measurement on a known arm. Do not compare this figure to numbers produced by a different harness.

Use with weight streaming

Pairs with xocialize/ltx-2.5-granules β€” a per-block re-layout of these exact bytes that streams the DiT from SSD instead of loading it resident. Output is bit-identical to the resident path (memcmp-gated, with a poisoned-slot negative control). Streaming replaces this file rather than supplementing it: fetch one or the other, not both.

Licence

LTX-2 Community License β€” see LICENSE.md and the acceptable-use policy snapshot in this repo. The weights are a derivative of Lightricks LTX-2.5 and carry its terms, including the revenue gate and non-compete in Attachment A. Redistribution is permitted under Β§3 with those terms propagated.

Downloads last month

-

Downloads are not tracked for this model. How to track
MLX
Hardware compatibility
Log In to add your hardware

Quantized

Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Model tree for mlx-community/ltx-2.5-mlx-ditq8

Finetuned
(17)
this model