A 22B Video-With-Sound Model That Runs on a 16GB Gaming GPU: LTX-2.5's Weights Are Out

TL;DR: Lightricks released LTX-2.5 open weights (Aug 11): one 22B model generates synchronized video and audio, 720p–4K, runnable locally from ~16GB VRAM with FP8 quantization.

If you have a 16GB consumer GPU, you can now generate video with synchronized audio from one local model. LTX-2.5's 22B checkpoints (full and 8-step distilled) are on Hugging Face, with ComfyUI and Diffusers support; FP8 quantization brings VRAM needs down roughly 40%, to about the 16GB floor. Native multishot holds character, lighting, and voice consistent across cuts — the thing that made local video toys fall apart at clip two. On top-end datacenter hardware Lightricks claims a 10-second clip in 6.8 seconds; at home you'll wait longer, but it runs.

One honest caveat: "open weights" here means the LTX-2.x Community License — free commercial use only under $10M revenue. For a weekend project, that's academic; know it before building a business on it.

Sources: VentureBeat, Open Source For You, DataNorth