Lightricks' LTX-2.5 Generates 10-Second Videos Faster Than Real Time

Runway just added LTX-2.5 to its platform, making the fastest open-weights video generation model available to anyone with a Runway account. The move is notable: LTX-2.5 is not a Runway-native model. It is an open-weights release from LTX, a company spun out of Israeli app maker Lightricks, and it is simultaneously available on Hugging Face, inside ComfyUI, and through the LTX API. Runway is simply one more on-ramp to a model that is already generating a lot of attention.

Faster than the clip itself

Self-hosted on 2x GB200 GPUs at 720p, LTX-2.5 generates a 10-second video clip in 6.8 seconds, which is faster than real time. That headline number needs context, but the competitive gap is still striking. The fastest closed alternatives, Omni Flash, Grok 1.5, and Veo 3.1, land at 52 to 70 seconds. Slower systems stretch much further: Seedance 2.0 at 196 seconds, FLUX 3 at 259, Seedance 2.5 at 317, and Kling 3.0 Pro at 398. The 6.8-second figure was measured self-hosted on two of NVIDIA's top-end GB200 chips at steady state, a configuration far beyond what most teams have in a rack. Still, even via the managed API, a 10-second clip takes 23.7 seconds via the LTX API. That is still 2-3x faster than the nearest closed competitor.

What actually changed under the hood

LTX rebuilt nearly every stage of the generation pipeline rather than bolting features onto an older core. The key architectural changes:

• New diffusion video decoder: Reduces visual artifacts in high-motion footage and reconstructs fine detail like text and faces, while preserving LTX's high compression ratio.
• Native multishot generation: A single generation can produce multiple connected shots, holding character, scene, lighting, visual style, and voice consistent across cuts. This is the feature that makes it usable for actual campaigns, not just demos.
• Gemma 4 language backbone: A custom Gemma 4 language backbone and dedicated prompt enhancer for more accurate handling of complex, multi-subject prompts.
• Diffusion Fidelity Rendering: A new video generation technology that allocates rendering compute by scene complexity, delivering industry-leading pixel quality that holds up frame by frame, even on a cinema screen.
• Physical AI checkpoint: A pretrained checkpoint tuned for physical AI and robotics, giving teams a base to fine-tune on domain data that looks nothing like cinematic video.

LTX-2.5 is built on a 22B-parameter asymmetric dual-stream diffusion transformer. The model comes in two API variants: ltx-2-5-fast for speed and low cost, and ltx-2-5-pro for higher fidelity. Fast produces portrait and landscape video up to 4K; Pro tops out at 1080p.

The open-weights angle matters more than it sounds

Most frontier video models are closed APIs. You send a prompt, get a clip back, and have no control over what happens in between. LTX-2.5 runs on any GPU with a minimum of 16GB of VRAM, deploys on-premises, at the edge, or via API, carries no visible watermark on output, and can be fine-tuned on a customer's own data and IP. That last point is what separates it from most competitors: you can train it on your brand's visual style, your product catalog, or your character designs, and keep everything on your own hardware.

The model was fine-tuned on 1M reasoning videos to improve multi-object motion, physical interactions, and prompt accuracy. LTX says its model family has now surpassed 33 million downloads, which the company claims makes it the most-used open world model line on the market.

Where it shines, and where it does not

LTX-2.5 is built for speed, consistency, and local deployment. The multishot generation and character consistency fixes are real improvements over earlier open models. The model produces natural-looking movement with reduced flicker and better frame consistency across the entire video. It can generate high-resolution HDR footage built for professional finishing.

The caveats are worth noting. The 6.8-second benchmark requires dual GB200 chips, hardware that most teams do not have. LTX claims roughly one-eighth the cost and one-seventh the render time of comparable models, on hardware ranging from data center GPUs down to a Mac, with a minimum of 16GB of VRAM. On a consumer RTX card, generation will be slower. The Pro API variant also caps at 1080p, so true 4K output requires the Fast variant, which trades some fidelity. And if you have existing LoRAs from LTX-2.3, they will need to be retrained for the new latent space.

Who should actually use this

The practical use-cases break down cleanly:

• Ad and social teams who need to batch-generate variations of the same creative brief, test multiple hooks, and refresh content before ad fatigue sets in. Local generation means no per-clip fees.
• Studios doing pre-visualization where multishot consistency lets you block out a full sequence before committing to a shoot.
• Developers building video pipelines who want to self-host, fine-tune on proprietary data, and avoid vendor lock-in. The model arrives with native integration into ComfyUI, the node-based workflow environment that has become the standard prototyping tool for anyone serious about open generative media.
• Robotics and physical AI teams who can use the dedicated physical AI checkpoint as a fine-tuning base for simulation and perception tasks.
• Researchers and students who want a fully open, commercially-licensed model to benchmark against or build on top of.

Access and pricing

On Runway, LTX-2.5 is available now to all subscribers under Runway's existing credit-based pricing. All plans include access to all models. For direct API access, the model is available as open weights on Hugging Face, inside ComfyUI, and through the LTX API. It is free to use for organizations under $10 million in annual recurring revenue; larger companies negotiate a license. The open-weights path via GitHub and Hugging Face costs nothing per generation once your hardware is set up.

The broader signal here is about where the industry is heading. The release anchors NVIDIA's month-long local AI series, launched the same day as its open Nemotron 3.5 Lightning agent model. The signal from both: open models, accelerated locally, are becoming default production infrastructure. Runway adding LTX-2.5 to its platform is a bet that the best video model for a given task might not be the one a platform built itself, and that giving users access to the open-weights ecosystem is now table stakes.