LTX-2 - open weight 4K/50fps video with audio from Lightricks

TG AI News·January 7, 2026 at 5:27 PM·
Trusted Source
Related tools:
Hugging Face
Lightricks, the company behind one of the first "content factories" LTX-Studio even before these factories flooded Twitter, made an interesting pivot. They released the video model **LTX-2** as open source, the first version of which was the basis of their reactor. The model ranks 23rd in the LM video arena, but that is not the main point. LTX-2 is the first **fully open** model capable of generating **native 4K video at 50 FPS** with synchronized audio (dialogues, music, SFX) lasting up to 20 seconds. At the core of LTX-2 is a **single asymmetric dual-stream transformer** for joint audio and video generation through cross-attention. The model is designed for consumer GPUs with 19B parameters (14 for video and 5 for audio). The weights, as well as **inference pipelines and training code**, are available as open source. Additionally, LTX-2 is quantized in NVFP8 (30% smaller, up to 2 times faster) and optimized for the NVIDIA ecosystem, with **ComfyUI** supporting it from day one. It is not entirely clear how this release fits with their main business model. While their ambitious desire to create their own model was understandable, the reason for releasing it as open source is quite unclear. After all, the experts from ComfyUI have already replicated the same LTX Studio in their Comfy and n8n on other models. UPD: The website mentions native 4K, but as correctly pointed out in the comments, it is actually an upscale.
LTX-2 - open weight 4K/50fps video with audio from Lightricks | AI News | AIventa