LTX, the open world-model company spun out of Lightricks, released LTX-2.5 on August 11, the newest version of its open-weights video and "world" model family — and it arrives with a headline number: a 10-second, 720p image-to-video clip rendered in 6.8 seconds, faster than real time, when self-hosted on two Nvidia GB200 superchips.
The speed claim comes with caveats. The 6.8-second figure was measured on top-end data-center hardware at steady state, a configuration far beyond most teams' reach; through LTX's own managed API the same job took 23.7 seconds at the higher 1080p resolution. Against rival APIs, LTX's end-to-end measurements put Google's Gemini Omni Flash at 52 seconds for the same task and MiniMax H3 at 180 seconds — roughly a sevenfold advantage for LTX-2.5, though the company notes the comparisons are its own.
Beyond speed, the release rebuilds much of the pipeline. LTX-2.5 introduces a new diffusion video decoder that reduces artifacts in high-motion footage and reconstructs fine detail like faces and text; native multishot generation that renders a full sequence as a single output, keeping characters, scenes and voices consistent across cuts; a custom Gemma 4 language backbone with a dedicated prompt enhancer; and a pretrained checkpoint tuned for physical AI and robotics workloads. An improved distilled model runs locally on Nvidia RTX GPUs with reduced memory requirements, and the model itself runs on any GPU with at least 16GB of VRAM.
The business model is a deliberate bet against closed APIs. LTX-2.5 is free for organizations under $10 million in annual recurring revenue, with larger companies negotiating a license; the LTX family has passed 33 million downloads, making it the most-used open world-model line on the market, according to the company. "Open weights is not some kind of one-time philanthropic fluke for us. It's the strategy," said CEO Zeev Farbman.
The "open weights" label, however, parts ways with open source in the fine print: the LTX-2.x Community License includes a revenue threshold and use restrictions, bans military and weapons development outright, and requires users to disclose machine-generated content and preserve embedded provenance features. On quality, LTX reported a 67% win rate in its own blind preference tests, but the independent Artificial Analysis arena — where Google's Gemini Omni Flash currently leads — has not yet scored LTX-2.5.
The model is available today on Hugging Face, natively in ComfyUI through a day-one partnership, and via the LTX API.
