Published: 12 August 2026
LTX released LTX-2.5 on 11 August 2026, an open weights world model for video generation, real time applications and physical AI. LTX-2.5 generates a 10 second 720p image to video clip in 6.8 seconds on two NVIDIA GB200 GPUs, adds native multishot generation that holds character, environment and voice consistent across cuts, and ships with a new diffusion video decoder and a Gemma 4 12B text encoder. The weights are available on Hugging Face and are free to use for organisations under 10 million dollars in annual recurring revenue.
What can LTX-2.5 do?
LTX-2.5 is a world model rather than a text to video tool in the narrow sense. Where a large language model predicts the next word, a world model predicts the next moment: it generates an environment, simulates how that environment behaves and lets a user or an agent act inside it. LTX positions LTX-2.5 for four groups: film and video studios, short form creators and advertising teams, developers of real time applications, and robotics teams. Asteria produces original film and video on LTX, Reactor runs LTX-2.5 on low latency infrastructure for interactive avatars and live worlds, and Markov Robotics uses LTX to develop how physical systems perceive and move.
The headline capability in this version is native multishot generation. LTX-2.5 renders a full sequence as a single output instead of stitching separate clips together, which keeps a character’s appearance, the scene, the lighting and the voice consistent from cut to cut. LTX-2.5 also adds automatic duration prediction, so the model generates the clip length that the requested action needs, native 4K HDR output, a native RAW workflow for professional colour and finishing pipelines, and a beta mode for precise editing of existing footage.
A separate pretrained checkpoint tuned for physical AI ships alongside the cinematic model. That checkpoint gives robotics teams a base to fine tune on domain data that does not look like film footage. LTX states a minimum requirement of 16GB of VRAM and says the model runs on any GPU, on premises, at the edge or through the API.
LTX-2.5 benchmarks and how it compares with Seedance 2.5, Veo 3.1 and Kling 3.0 Pro
All published figures for LTX-2.5 come from LTX’s own evaluations, and LTX labels the quality results as preliminary.
- In LTX’s image to video speed benchmark, a 10 second clip takes 6.8 seconds when self hosted on two NVIDIA GB200 GPUs at 720p and 23.7 seconds through the LTX API.
- The fastest closed alternatives in the same table, Omni Flash, Grok 1.5 and Veo 3.1, land between 52 and 70 seconds. Seedance 2.0 takes 196 seconds, FLUX 3 takes 259 seconds, Seedance 2.5 takes 317 seconds and Kling 3.0 Pro takes 398 seconds.
- On premises, that makes LTX-2.5 roughly 7.6 times faster than the nearest closed alternative and about 58 times faster than the slowest system in the comparison.
On visual quality, LTX ran an automated artifact score across 98 text to video prompts and 10 models, counting blotches, broken texture and smeared areas, where a lower score is cleaner.
LTX 2.5 Pro scores 0.28 and LTX 2.5 Fast scores 0.39, ahead of:
- Flux 3 at 0.45,
- MiniMax at 0.46,
- Wan 2.6 at 0.65,
- Seedance 2.5 at 0.69,
- the previous LTX 2.3 Pro at 0.74,
- Kling 3 Pro at 0.76,
- Google Veo 3.1 at 1.20.
LTX notes that this scoring is automated rather than human graded.
LTX also reports blind side by side testing in which evaluators saw two videos from the same prompt with model names hidden and randomised positions. LTX 2.5 wins 67 percent of those matchups, followed by Seedance 2.5 at 65 percent, Omni Flash at 55 percent, MiniMax H3 at 50 percent, Seedance 2.0 at 44 percent, Wan 2.6 at 42 percent and FLUX 3 at 28 percent. LTX has not published the evaluator pool size or the full prompt set, and no third party has reproduced either result, so these numbers should be read as vendor claims until independent testing appears.
What changed in the LTX-2.5 architecture
LTX rebuilt most of the generation pipeline rather than adding features to the LTX-2.3 core. The most visible change is a new diffusion video decoder that reduces artifacts in high motion scenes while keeping the high compression ratio the LTX family relies on. Prompt handling moved to a custom Gemma 4 12B text encoder paired with a dedicated prompt enhancer, which LTX credits for better comprehension of complex, multi subject prompts from shorter instructions.
The new generation technique is called Diffusion Fidelity Rendering. It builds motion and structure inside an 8 times temporally compressed latent space, then generates high fidelity keyframes that anchor visual detail. The number of keyframes adapts to how complex the scene is and to the available compute budget, which is how LTX allocates rendering cost by scene rather than uniformly across a clip.
Two further changes affect production cost. LTX shipped an improved distilled model trained on a larger dataset with reinforcement learning, which it says delivers near full model quality at significantly lower compute and with fewer retries. And LTX-2.5 includes a raw pretrained foundation checkpoint intended for fine tuning, so teams can adapt the model to their own data and domain instead of working only with the released production weights.
LTX-2.5 availability, pricing and licensing
LTX-2.5 became available on 11 August 2026 in three forms at once. The open weights are on Hugging Face under the Lightricks organisation, the model has day one support in ComfyUI through a launch partnership, and the LTX API serves teams that prefer managed infrastructure. Source code is on GitHub in the Lightricks/LTX-2 repository, with documentation at docs.ltx.video.
API pricing is billed per second of generated video: 0.09 dollars per second at 720p, 0.15 dollars per second at 1080p, 0.19 dollars per second at 2K and 0.37 dollars per second at 4K. The open weights carry a permissive licence with no mandatory branding requirement and are free to use for organisations under 10 million dollars in annual recurring revenue. Larger companies negotiate a separate licence with LTX.
LTX describes the LTX family as the most used open world model line, citing more than 33 million downloads. The release also anchors NVIDIA’s month long local AI series and landed the same day as NVIDIA’s own Nemotron 3.5 Lightning agent model, which is why much of the launch coverage frames LTX-2.5 in terms of local inference on NVIDIA RTX GPUs and DGX Spark rather than cloud capacity.
The model page, the benchmark tables and the licence terms are at the official LTX2.5 announcement by LTX, and the weights of LTX 2.5 can be found at at Huggingface.
