LTX-2.5 Launch: NVIDIA-Accelerated Open Weights World Model for Local Video Production
LTX
NVIDIA
Google/DeepMind
xAI
LTX has released LTX-2.5, an open weights world model for video generation optimized for local inference on NVIDIA RTX GPUs and DGX Spark. The model enables fast, consistent multishot video generation on a single desktop, with on-prem speed significantly faster than closed alternatives. The release aligns with NVIDIA's local AI series, which also introduced the Nemotron 3.5 Lightning agent model.
LTX launched LTX-2.5, an open weights world model for video generation, real-time applications, and physical AI, optimized to run locally on NVIDIA RTX GPUs and DGX Spark, reducing VRAM requirements. It features native multishot generation for character consistency, a new decoder reducing artifacts, and a Gemma 4 language backbone. Benchmarks show a 10-second clip generated in 6.8 seconds on 2x NVIDIA GB200, compared to 52-70 seconds for fastest closed alternatives and up to 398 seconds for slower ones. NVIDIA's local AI series also includes Nemotron 3.5 Lightning, an open 30B mixture-of-experts model, and NeMo Switchyard library. LTX-2.5 is available as open weights on Hugging Face, natively in ComfyUI, and via LTX API, free for organizations under $10M annual recurring revenue. The model family has over 33 million downloads, with applications cited in film studios like Asteria, real-time platforms like Reactor, and robotics with Markov Robotics.
- Abbreviations
- GPU = Graphics Processing Unit — графический процессор
- VRAM = Video Random Access Memory — видеопамять
- API = Application Programming Interface — программный интерфейс приложения
- A/B = A/B testing — A/B-тестирование
- MoE = Mixture of Experts — смесь экспертов
- LLM = Large Language Model — большая языковая модель
- IP = Intellectual Property — интеллектуальная собственность
- ARR = Annual Recurring Revenue — годовой повторяющийся доход
Source: MarkTechPost —
original