LTX, a spin-off of Israeli company Lightricks, released LTX-2.5 on August 11, 2026 — a fully redesigned open video generation model that generates a 10-second 720p clip in 6.8 seconds on two NVIDIA GB200 GPUs, outpacing closed competitors from Google, xAI, and OpenAI in both speed and quality.
What Happened
LTX announced LTX-2.5 with three key architectural updates. The diffusion video decoder has been completely redesigned: it reduces artifacts during high scene dynamics and restores fine details, including text and faces. The model now features native multishot generation — the ability to create connected shots while preserving character, environment, and voice without post-processing or stitching. The language core has been replaced with a custom Gemma 4 featuring a prompt-enhancer for better understanding of complex prompts. API pricing: $0.09 per second for 720p, $0.15 for 1080p, $0.19 for 2K, $0.37 for 4K. Weights are published on Hugging Face, and the license is free for organizations with revenue up to $10 million ARR. ComfyUI support has been integrated from day one.
Context
The LTX model family has accumulated over 33 million downloads to date. LTX is a spin-off of Lightricks, the Israeli developer of Facetune. Previous versions of LTX have already demonstrated that open-weights video generation models can compete with closed solutions in terms of quality. The AI video industry has been dominated by proprietary APIs from major cloud providers — Google with Veo and Gemini Omni Flash, xAI with Grok, and OpenAI with Sora. LTX-2.5 continues the open-weights strategy established by Stable Diffusion for images, but scales it to video.
Why This Matters for the Industry
LTX-2.5 is the first open-weights video generator with a speed of 6.8 seconds for a 10-second clip, which is significantly faster than closed competitors: Gemini Omni Flash generates 8 seconds in 52 seconds, Grok 1.5 in 63 seconds, and Veo 3.1 in 70 seconds. In blind human tests, LTX-2.5 wins 67% of the time, outperforming Seedance 2.5 (65%) and Gemini Omni Flash (55%). The combination of open weights, a license up to $10 million ARR, and native ComfyUI support creates a channel for open-weights models to penetrate production pipelines of studios and corporations. ComfyUI integration accelerates the emergence of custom pipelines, LoRA adaptations, and specialized products. The 6.8-second speed indicates sublinear generation time and an efficient scheduler, which could become a pattern for the next generation of video models.
Why This Matters for Users
The model can be run locally on a GPU with 16 GB of VRAM. Organizations with annual revenue under $10 million can use LTX-2.5 for free. Users get one of the fastest open video generators on the market — generation of connected scenes, experiments with physical AI and robotics without sending data to the cloud and without forced watermarks. Weights are available on Hugging Face, and ComfyUI nodes are already working. API integration allows embedding generation into your own products at a price starting from $0.09 per second.
What Is Still Unknown / Limitations
The exact architecture of the new diffusion video decoder and the parameters of the custom Gemma 4 have not been disclosed. The model's parameter count, training dataset composition, and fine-tuning methodology (including the RLHF pipeline) have not been published. The methodology of the blind tests that yielded a 67% win rate — sample size, demographic diversity, statistical significance — has not been disclosed. Exact VRAM requirements for resolutions above 720p are not specified. Independent benchmarks have not yet been conducted.
Sources
Author
Look at AI, editorial team