LTX-2 mixture-of-experts variant cuts inference cost while dense model scales quality
Lightricks CEO Zeev Farbman outlined the next LTX release, which ships both a dense model and a mixture-of-experts variant, plus open training recipes for domain-specific LoRA fine-tunes.
LTX-2, Lightricks' next-generation open video synthesis model, will ship in two architectural flavors: a traditional dense checkpoint and a mixture-of-experts (MoE) variant that activates only the network layers needed for each frame. CEO Zeev Farbman detailed the roadmap this week, explaining that the MoE approach scales capability and quality without linear compute growth—a fundamental shift that changes what the model can do at a given inference cost. Both versions get a significantly upgraded text encoder designed to parse multi-scene prompts that earlier LTX releases simplified or ignored, improving understanding of complex written instructions with multiple actions or camera moves layered into a single prompt. The release also includes new attention kernels and improved low-precision support, widening the range of consumer hardware that can run the model locally.
Lightricks is opening more of its training stack alongside the weights. The company will release trainer recipes and LoRA tooling so practitioners can build domain-specific variants—human motion, product visualization, architectural walkthroughs—on top of the base checkpoint. Enterprise customers get a post-training customization layer for closed-data fine-tuning without retraining from scratch. Farbman described a three-tier picture: foundation weights, open LoRA recipes for the community, and private fine-tuning for teams with proprietary datasets. The move positions LTX's MoE variant as a direct competitor to dense architectures like Wan, Hunyuan, and Kuaishou, offering a speed-quality tradeoff that could matter for real-time or batch workflows. No release date was announced, but Farbman said the work represents "what we're building right now and what you'll see soon."
