Lightricks released LTX-2.3 on 5 March, the latest update to its open-source video generation model family. The release is available on Hugging Face and via the LTX API.
The headline additions are a redesigned VAE architecture producing sharper textures and cleaner edges, a larger text connector that the company says handles complex multi-subject prompts more accurately, and native portrait video generation up to 1080×1920 — trained on vertical data rather than cropped from landscape output. Audio quality has also been improved, with silence gaps and noise artifacts filtered from the training set.
The portrait video support is the most commercially pointed addition. Vertical format is the default delivery for short-form platforms, and until now most models have handled it as an afterthought. Building native 9:16 generation into the model rather than bolting it on post-hoc matters for production workflows, particularly for agencies and creators working at volume.
LTX-2.3 is available as open weights, free for commercial use for companies with under $10 million in annual revenue. Organisations above that threshold require a commercial licence. The model also ships with a local desktop video editor and MCP integration for developers building generation into their own pipelines.
Lightricks released the full LTX-2 model with open weights in January. The 2.3 update continues a release cadence that has positioned the company as the most active open-source competitor in a market otherwise dominated by closed proprietary systems.
