Five days after Wan 3.0 went API-only and nine after MiniMax-H3 shipped weights that excluded the US, EU, UK, and Korea, LTX released 2.5 as the counter-argument: 22B parameters on Hugging Face, day-one ComfyUI support, free for commercial and production use under $10M ARR, no mandatory branding, and no geographic carve-outs. The company spun out of Lightricks — the Jerusalem outfit behind Facetune and Videoleap — earlier this year to become an open world-model company, and says the LTX family has passed 33 million downloads. The headline is speed. A 10-second 720p image-to-video clip renders in 6.8 seconds self-hosted on two GB200s — faster than real time, and in LTX's own benchmark table roughly 10× quicker than Veo 3.1 and 47× quicker than Seedance 2.5. That comes from Diffusion Fidelity Rendering, which builds motion and structure in an 8× temporally compressed latent space while generating high-fidelity keyframes alongside it, spending compute on the demanding moments and skipping it elsewhere, then resolving the whole thing in a dedicated diffusion pass. A new diffusion video decoder replaces the old VAE reconstruction, which is why faces stop softening under compression. Minimum VRAM is 16 GB; it runs on RTX cards, DGX Spark, and Mac, with a distilled variant tuned alongside NVIDIA. The features aimed at actual production are the more interesting part. Native multishot generates several connected shots as one output, holding character, environment, lighting, and voice across the cuts — the consistency problem solved inside the model rather than in the edit. A custom Gemma 4 12B text encoder with a prompt enhancer reads multi-subject direction properly; a duration predictor picks the clip length the action needs. And LTX-2.5 takes 16-bit linear EXR in ACES in and returns ACES out with the range intact, plus native 4K HDR and a RAW workflow — unglamorous plumbing that means a generated shot can pass through a professional finishing pipeline instead of dead-ending as an MP4. Paul Trillo, whose studio Asteria is a launch partner, singles out exactly that. Two caveats: the speed and artifact numbers are LTX's own, self-run and labeled preliminary; and the managed API is billed per second — $0.09 at 720p up to $0.37 at 4K — for anyone not running it themselves.