LTX-2.5 system requirements: VRAM, RAM and the size of every file

Lightricks' documentation gives one set of numbers for running LTX-2.5 on your own machine. The minimum is an NVIDIA GPU with 32 GB of VRAM (“more is better”), 32 GB of system RAM, 100 GB of free storage, CUDA 12.7 or newer and Python 3.12 or newer. The recommended setup is an A100 with 80 GB or an H100, 64 GB of RAM or more, and 200 GB or more on an SSD.

Why the numbers are that high

The model is a 22-billion-parameter transformer, and in bf16 its file alone weighs 42.02 GB, more than the 32 GB minimum. The text encoder that reads your prompt, a Gemma 4 12B with LTX's projections, is another 26.26 GB in bf16. At full precision the transformer alone is larger than a 32 GB card, and the two together come to about 68 GB, which is why the routes below for cards near the minimum either compress the weights or move part of them to system RAM.

The size of every file

From the file listing of the Hugging Face repository, as checked on 6 November. Transformers: the dev and the distilled checkpoints are 42.02 GB each in bf16 and 21.50 GB each in ComfyUI's int8 format; the distilled one in NVFP4 is 18.72 GB. Text encoder: 26.26 GB in bf16, 15.37 GB in int8. The two video VAEs are 1.47 GB (the heavier, higher-quality DiffVAE) and 1.45 GB (the faster Conv VAE); the audio VAE is 0.36 GB, the ×2 spatial upscaler 1.00 GB, the ×2 temporal upscaler 0.26 GB, the distilled LoRA for dev workflows 8.90 GB, and the duration head, which picks a clip length from the prompt, about 4 MB.

Three sets to download, not the whole repository

The repository holds about 201 GB in all, close to the 200 GB the documentation recommends, but nobody needs all of it. It is gated: you sign in to Hugging Face and accept the LTX-2.x Community License on the model page before any file downloads, which is why the card's download command starts with hf auth login. The bf16 distilled set that the model card's own download command fetches (transformer, text encoder, video and audio VAEs, duration head, spatial upscaler) comes to about 71 GB. The ComfyUI int8 set, with the int8 transformer and int8 text encoder in their place, is about 40 GB. In ComfyUI with the NVFP4 transformer instead of the int8 one it is about 37 GB.

Which set for which GPU

The int8 files are ComfyUI only: the model card says Lightricks' own PyTorch package, ltx-pipelines, does not load them. The NVFP4 transformer works in ComfyUI or in ltx-pipelines with --quantization nvfp4-prequant, and for that ltx-pipelines route the card names Blackwell GPUs and the ltx-kernels package, so check your card's generation before you download it. For ltx-pipelines the card's low-VRAM advice is to keep the bf16 files and add --quantization fp8-cast --offload cpu: the bf16 transformer is downcast to fp8 on the fly and parts of the model are offloaded to the CPU, into system RAM.

In ComfyUI

Official LTX-2.5 workflow templates ship with ComfyUI, which has run the model since version 0.32. There the int8 transformer and int8 text encoder are the smaller pair to start from on a card at the 32 GB minimum, and the model card describes the Conv VAE as faster and lighter than the higher-quality DiffVAE. The Diffusers pack, Lightricks/LTX-2.5-Diffusers, is a third route; LTX-2.5 support is not yet in a Diffusers release, so the card installs Diffusers from its main branch, and its example turns on model CPU offload and VAE tiling, noting that the second stage decodes at twice the size.

Smaller community builds

Third-party GGUF conversions of the distilled model on Hugging Face shrink the transformer further: one such repository lists files from 12.65 GB at Q3_K_S to 23.60 GB at Q8_0. They are not Lightricks' releases and Lightricks publishes no requirements for them, so treat their VRAM claims as their authors', and expect quality to fall with the bit count.

Without a GPU of your own

Lightricks runs LTX-2.5 in its API Playground at console.ltx.video, which needs no download at all, and ComfyUI has partner nodes for the hosted version. For comparing it with the other open model that generates sound with the picture, see LTX-2.5 against MiniMax H3.