One open world model. Fine-tune it into whatever you're building.
LTX-2.5 ships two ways: a finetuned checkpoint built for video production and real-time avatars, and a raw, pretrained base for teams building robotics, autonomous vehicles, and other physical AI applications. Same 22B foundation. Open weights. Self-hosted on infrastructure you control.
.svg)
Key Capabilities
Start from the pretrained foundation
A raw, non-SFT base built before finetuning. Adapt it to your domain instead of one already locked into cinematic output.
Take it beyond video
The same foundation behind multishot video can be redirected toward robotics, autonomous vehicles, or drones.
Fast enough for production
Self-hosted on 2x NVIDIA GB200 GPUs, LTX-2.5 generates a 10-second 720p clip in 6.8 seconds, faster than real time.
Robotics and manipulation
Finetune the pretrained checkpoint on egocentric footage and action data to model how a robot's actions change what it perceives next. Build and evaluate on infrastructure you control, using data that stays in your environment.
Autonomous vehicle and drone simulation
Generate synthetic scenarios, rare weather, near-misses, and edge cases that are too costly or dangerous to capture repeatedly on real roads or in flight, without putting a vehicle or drone at risk.
Industrial and logistics automation
Model how a facility or process behaves and evolves over time, then test changes against that model before touching the physical environment.
Real-time and interactive applications
Launch partner Reactor runs LTX-2.5 on low-latency infrastructure to power interactive avatars, live worlds, and real-time workloads on the same base model.
How it works
Input
Technical characteristics:
- Conditioning — Text, image, or video conditioning for the main LTX-2.5 checkpoint (general-purpose generation).
- Domain data — Domain-specific action or state data — robot trajectories, sensor logs, simulation states — for fine-tuning the LTX-2.5 pretrained checkpoint.
- Fine-tuning — Runs through the open source LTX Trainer, paired with Google's Gemma 4 text encoder shipped alongside the checkpoint.
Output
Technical characteristics:
- Main checkpoint — Connected multishot video with consistent character, environment, lighting, voice, and style across cuts.
- Fine-tuned deployments — Predicted or generated future states of an environment, action outcomes, simulated scenarios, or scene evolution, depending on the fine-tuning target.
What users are building with LTX
FAQs
What is a world model, and is LTX-2.5 one?
A world model represents an environment and predicts how it's likely to change over time, rather than only describing what's already there. LTX-2.5's main checkpoint models motion and continuity across connected shots for production video, and the same 22B architecture ships as a pretrained checkpoint for teams extending it into non-cinematic domains.
What's the difference between the main LTX-2.5 checkpoint and the Pre-Trained checkpoint?
The main release is supervised fine-tuned and post-trained for general-purpose video and audio generation. LTX-2.5 Pre-Trained is the raw base from before that stage, intended for researchers and teams doing aggressive fine-tuning into a new domain rather than general video generation out of the box.
Can I fine-tune LTX-2.5 for robotics or autonomous vehicles?
Yes, that's the intended use of the Pre-Trained checkpoint: egocentric robotics and manipulation, action-conditioned world prediction, autonomous vehicle and drone data, or industrial digital twins, via the open source LTX Trainer.
These applications require domain-specific fine-tuning and evaluation. The pretrained checkpoint is a foundation to build from, not an out-of-the-box robotics or autonomous-driving system.
How fast does LTX-2.5 run in production?
Self-hosted on 2x NVIDIA GB200 GPUs, LTX-2.5 generates a 10-second 720p clip in 6.8 seconds, faster than real time. Actual speed depends on hardware, resolution, and whether you're running the main checkpoint or a fine-tuned derivative.
How do I run it, and what does it cost?
LTX-2.5 is available as open weights on Hugging Face. Fine-tune with the LTX Trainer and self-host on your own GPUs through ComfyUI or the open source pipeline, down to a single Mac for local iteration. Or call the main model through the LTX API if you'd rather not manage infrastructure.
Commercial use is free under the LTX Community License for organizations under $10M in annual revenue. Above that, or if you're embedding LTX in a product you sell, a commercial license applies.
Who is already building physical AI on LTX?
Markov Robotics uses LTX to develop how physical systems perceive and move through the world. Reactor runs LTX-2.5 on its low-latency infrastructure for interactive avatars, live worlds, and real-time workloads.
