The Video Production Stack Now Fits on One Desk: LTX-2.5 Launches as NVIDIA-Accelerated Open Weights World Model

Details

Lightricks has released LTX-2.5, an open weights video world model built for local inference on consumer NVIDIA RTX GPUs. The model generates a 10-second clip in 6.8 seconds on 2×NVIDIA GB200 hardware on-premise, versus 23.7 seconds via the LTX API. Published benchmarks show it significantly outpacing closed alternatives: Veo 3.1 takes 52-70 seconds, Seedance 2.0 takes 196 seconds, and Kling 3.0 Pro stretches to 390+ seconds.\n\nKey capabilities include native multishot generation — rendering a coherent multi-shot sequence as one piece, maintaining character consistency across cuts. An updated decoder reduces motion artifacts in high-action scenes. The Gemma 4 language backbone improves text-following fidelity. LoRA fine-tuning is supported for branding and character locking. The model runs locally in ComfyUI.\n\nThe commercial case: ad teams face 7-10 day creative fatigue cycles, and local generation with no per-clip fees enables high-volume iteration — a week of content variations can run overnight on a consumer GPU. Small studios can match the output volume of larger production setups.\n\nNVIDIA is anchoring LTX-2.5 as part of a broader local AI push, released the same week as Nemotron 3.5 Lightning, signaling open-weights locally-run generative media is becoming production infrastructure.