Lightricks has open sourced LTX-2, the first production-ready audio-video generation model with fully open weights, enabling creators and enterprises to locally generate native 4K video with synchronised sound, without cloud lock-in or black boxes.
Lightricks has announced the open source release of LTX-2, positioning it as the first production-ready audio and video generation model to combine truly open weights with native 4K output and fully synchronised sound. The release includes complete access to model weights, inference pipelines, and training code, marking a significant shift in open-source generative media.
LTX-2 is the first open model capable of natively generating synchronised audio and video, delivering expressive sound, accurate lip sync, and high audio fidelity. Designed for real-world production workflows, the model generates up to 20 seconds of audio-video content at native 4K resolution, running at 50 frames per second.
Built for on-device deployment, LTX-2 runs locally on consumer GPUs, eliminating dependence on cloud-only systems while ensuring privacy, IP protection, and regulatory compliance. The release includes both a full model and a distilled variant, enabling faster inference with minimal quality trade-offs and allowing teams to balance performance and fidelity based on available hardware.
Production-grade optimisations are included out of the box. LTX-2 has been quantised to NVFP8, reducing model size by approximately 30 per cent and improving performance by up to 2×. ComfyUI has been optimised specifically for LTX-2, further lowering compute requirements compared to existing open-source audio-video models.
Optimised for the NVIDIA ecosystem, LTX-2 supports GeForce RTX GPUs and scales from local PCs to NVIDIA DGX Spark and enterprise data centres. The open release also enables deep customisation, allowing teams to train proprietary IP, embed custom visual and sonic styles, and create LoRA fine-tunes, while supporting transparent and reproducible multimodal research through released benchmarks and training insights.
“LTX-2 is the first truly open audio-video model, released with open weights and training code, and designed to run locally on consumer GPUs. It delivers the kind of quality and performance teams usually associate with closed systems, without giving up control, transparency, or the ability to customize. We believe this release marks a meaningful shift for both research and real-world production pipelines, expanding what teams can build and who gets to build it,” said Zeev Farbman, Co-founder and CEO of Lightricks.
LTX-2 is available via a self-serve API, the LTX platform, and integrations including Fal, Replicate, ComfyUI, and OpenArt. The model is free for academic use and for commercial use by companies with under $10 million in ARR, while larger organisations can access the same open-weight model through a commercial licence with enterprise-grade support.
