Skip to main content
Back

Lightricks Ships LTX-2.5, a 22B Open Video Model With Native Audio

Lightricks drops a 22B open-weights audio-video model that renders a 10-second 720p clip in under seven seconds and keeps character identity across multiple shots.

Model, Full Access
  • Lightricks released LTX-2.5, a 22B open-weights audio-video diffusion transformer with day-one ComfyUI support.
  • Native multishot generation holds character, lighting, and voice consistent across cuts in one pass.
  • A new diffusion video decoder replaces the VAE stage for sharper faces, textures, and on-screen text.
  • Uses a Gemma 4 12B text encoder and ships bf16, int8-convrot, and NVFP4 quantized checkpoints.
  • Free commercial use under $10M ARR; larger companies negotiate a paid license.
  • Generates a 10-second 720p clip in 6.8 seconds on dual GB200 superchips; roughly 1.5 minutes on RTX 5090.

Lightricks just pushed LTX-2.5 onto Hugging Face, and it is the most feature-complete open release the LTX line has shipped so far. The 22B-parameter open-weights audio-video foundation model introduces native multishot scenes, auto duration, and 4K HDR output, with official ComfyUI workflow templates available on day one. Weights, quantized variants, LoRAs, upscalers, VAEs, and text encoders are all downloadable, and the license is unusually generous for a model of this scale.

What actually shipped

LTX is the open world model company spun out of Lightricks, and 2.5 arrives natively integrated into ComfyUI, the node-based tool that has become the de facto prototyping environment for open generative media. The model is available immediately in three forms: as open weights on Hugging Face, inside ComfyUI through a launch partnership, and via the LTX API for teams that prefer managed infrastructure.

Licensing is free for organizations under $10 million in annual recurring revenue; larger companies negotiate a paid arrangement. Revenue is measured across the whole entity including subsidiaries, and transferring fine-tunes may require a paid license. Small teams can ship commercial products without per-generation billing, while Lightricks retains a monetization path for anyone at scale.

The Hugging Face repo ships a split, ComfyUI-aligned pack rather than a monolithic checkpoint:

  • A distilled 22B DiT (bf16) with a fixed 8-step schedule and CFG=1
  • A full, trainable dev DiT (bf16)
  • ComfyUI-only int8 quantizations with convolutional rotation
  • An NVFP4 distilled build for Blackwell GPUs
  • A Gemma 4 12B text encoder with learned projections
  • Two video VAEs (a heavier diffusion VAE and a faster conv VAE), plus an audio VAE with vocoder
  • x2 spatial and x2 temporal latent upscalers, a duration head, and a distilled LoRA

The speed number, with asterisks