LTX-2.5: Lightricks takes generative video weights outside the API, but the licence stays “conditional”

Lightricks (LTX) released LTX‑2.5 on 11 August 2026, a generative video‑audio model with open weights; the press note carrying partner statements circulated on 13 August. The official Hugging Face model card describes a 22‑billion‑parameter diffusion transformer, called “asymmetric dual‑stream” in the ComfyUI release note. Several weight variants are published: a distilled DiT (bf16, int8, NVFP4) and a full, trainable DiT (bf16, int8), together with video and audio VAEs, an upscaler, LoRAs and a “duration head” patch. The text encoder is a 12‑billion‑parameter Gemma 4 tuned for LTX; the official GitHub repository states that standard Gemma 4 is not compatible.

The announced features include native multishot generation in a single pass (character, environment and voice stay consistent across cuts), audio generated in sync with the video, automatic duration prediction, 4K HDR output and “Diffusion Fidelity Rendering”, which allocates compute according to scene complexity instead of applying a fixed compression rate. The official repository lists the generation modes: text→video, image→video, video→video, audio→video, keyframe interpolation and region regeneration. Software requirements from the model card: Python ≥ 3.12, CUDA ≥ 12.7, PyTorch ~2.7; the distilled models use a fixed 8‑step schedule. ComfyUI supports it from day one, with official templates for text→video, image→video and first/last frame→video.

On the practical side, Lightricks says a 10‑second clip generated from an image takes 6.8 seconds on two NVIDIA GB200 superchips and 23.7 seconds through the LTX API; the stated minimum is 16 GB of VRAM. Published API pricing is $0.09/s at 720p, $0.15/s at 1080p, $0.19/s at 2K and $0.37/s at 4K. In the figures LTX circulated, the artefact score (lower is better) is 0.28 for LTX‑2.5 Pro and 0.39 for LTX‑2.5 Fast, against 0.69 for Seedance 2.5 and 1.20 for Google Veo 3.1; LTX calls it preliminary.

The licence is the LTX‑2.x Community License: commercial use is free below $10 million in annual revenue, with a paid agreement above that threshold. It is not an OSI‑approved open source licence, and it imposes obligations to notify modifications, plus a ban on removing or circumventing synthetic‑content disclosure mechanisms. The release also includes a pre‑trained checkpoint for physical AI and robotics, meant for fine‑tuning on domain data. On the positioning as a “world model”, LTX/Lightricks co‑founder and chief executive Zeev Farbman says in the notes circulated on 13 August: “World models face challenges that LLMs never had to solve, like holding motion, space, and sound consistent across time”. On local execution, NVIDIA’s senior director for Local AI, Gerardo Delgado Cabrera, says in the same 13 August notes: “Local models enable creators and developers to freely explore their ideas thanks to their local GPUs”.

None of these numbers has so far been reproduced by third parties: they are all vendor claims. The comparison LTX published is methodologically uneven: rivals are measured mostly at 720p against an internal LTX measurement at 1080p; the rivals’ timings include queueing and third‑party hosting (fal.run); the Veo 3.1 comparison uses an 8‑second clip against 10 seconds for LTX. The 6.8 seconds refer to two GB200s, hardware few people have; there are no independent measurements at the 16 GB VRAM minimum, and no checks on the degradation caused by int8/NVFP4 quantisation, fp8 casting and CPU offloading. Multishot consistency, the advantage of the diffusion decoder over the VAE and the behaviour of the duration predictor all remain unverified. The artefact score is an automatic measurement over 98 prompts run through 10 models, and LTX has not published the full prompt set. The model card itself lists explicit limits: the model is not designed to provide factual information, it can amplify social biases, and prompt adherence varies with writing style, up to generations that do not match the request. Minor note: one field on the Hugging Face page carries the date 6 January 2026, inconsistent with the 11 August announcement (most likely a leftover from LTX‑2). The 33+ million downloads figure refers to the whole LTX family, not to this model.

LTX opens a useful path: downloadable weights, pipelines ready in ComfyUI, a concrete nod to robotics. But the revenue‑linked licence condition is a reminder that “open” here mainly means inspectable, not free of strings. The next interesting news will not be the demos, but independent reproductions and the first meaningful benchmarks on ordinary GPUs. — Olya