LTX-2.5 ships as a 22B open-weight video model with audio

Lightricks has shipped LTX-2.5, a 22B-parameter open-weight model that generates video and synchronized audio together from text, image or video prompts. The weights and the full model card are on Hugging Face, behind a click-through access gate.

The headline change over LTX-2 is native multishot generation. Where earlier LTX models produced a single continuous take, LTX-2.5 turns one prompt into a sequence of connected shots, and the card says character identity, environment, lighting, voice and visual style hold across the cuts. ThursdAI lists it as a 22B open-weights video model with multi-shot support.

Four more things changed. A new diffusion video decoder replaces the VAE reconstruction stage. A custom Gemma 4 12B text encoder holds longer prompts together. An optional duration predictor sets the frame count from the prompt, and a prompt enhancer expands a short line into a longer cinematic instruction. Lightricks also says the distilled checkpoint is substantially improved.

LTX-2.5 ships as split safetensors files, one per component. The distilled transformer runs a fixed 8-step schedule with CFG set to 1; the dev transformer is the full, trainable version. Lightricks says most LoRAs and IC-LoRAs trained on LTX-2.3 run on LTX-2.5 without changes.

Audio is generated jointly through an audio VAE and vocoder, in the same pass as the video.

One caution the model card states plainly: it publishes no benchmark scores. Nothing on the card ranks LTX-2.5 against another model.

The page does not state a release date. It showed more than 9,200 likes and 2.1 million downloads in the last month, read on October 25, 2026, and it sat at number six on Hugging Face's trending models list that day, so the weights may have been circulating before they started climbing.

You can run LTX-2.5 through the ltx-pipelines Python package, official ComfyUI templates, or a Diffusers pack installed from main. It wants Python 3.12 or newer, CUDA 12.7 or newer and PyTorch around 2.7; the low-VRAM path is --quantization fp8-cast --offload cpu. Quantized builds come in bf16, Comfy int8 and NVFP4, and the NVFP4 file needs a Blackwell GPU.

The license is the LTX-2.x Community License. Commercial and production use is free for entities under $10 million in annual revenue; above that you need a paid commercial agreement, and transferring fine-tunes may need a paid license too. This is open weights, not open source in the OSI sense.

ThursdAI reports the model generates 10 seconds of 1080p in 23.7 seconds on fal and needs a minimum of 16GB of VRAM, figures that do not appear on the model card.

Request access on the model page, pull the distilled checkpoint first, and start from the ComfyUI template.