LTX-2.5: open-weights video+audio, 10-sec clips, runs on 12GB VRAM

Lightricks' 22B audio-video model is free to self-host under $10M revenue, does up to 20-second 4K with sound, and ships to ComfyUI, diffusers, and a $0.09/sec API.

Nowline SEP 2 8:00 AM banner

Top AI stories from the last hour

Top AI stories from the last hour

Copy markdown

  • One 12GB GPU is the whole studio

    The distilled checkpoint runs locally on a single 12GB card with CPU offload and tiling: image-to-video, text-to-video, and synchronized audio, no cloud account required. On two NVIDIA GB200s it renders a 10-second clip in 6.8 seconds; most builders will run it slower but for free.

  • Grab it three ways

    Open weights sit on Hugging Face (Lightricks/LTX-2.5 plus an LTX-2.5-Diffusers pack), with a day-one ComfyUI workflow and a diffusers LTX2Pipeline for code. Prefer not to host? The managed API at console.ltx.io does the same generations.

  • Undercuts the closed video APIs

    Hosted pricing is $0.09 per second on the Fast tier: $0.90 for a 10-second clip, up to 20 seconds and 4K per generation, with the LTX family now past 33M downloads. Self-hosting and fine-tuning stay free if your company clears under $10M ARR.

  • Read the license before you ship

    Above $10M ARR you must negotiate a commercial agreement, training a competing video model is barred, and outputs must be disclosed as machine-generated. Fine for indie and small-team projects; check the terms before you put it in production.