LTX-2.5 Open-Weight AI Model Generates 10-Second Videos in Just 6.8 Seconds on Nvidia Superchips

LTX-2.5 Delivers Unprecedented Speed in Video Generation

LTX released LTX-2.5 on Tuesday, marking a significant advancement in open-weight AI model capabilities for video generation and robotics applications. The model generates a 10-second, 720p video from an image in just 6.8 seconds on two Nvidia GB200 superchips, running faster than real-time2

2

. Through LTX's managed API, the same task takes 23.7 seconds at 1080p resolution. By comparison, competitors require substantially longer processing times: Google's Gemini Omni Flash needs 52 seconds, xAI's Grok 1.5 takes 63 seconds, Google's Veo 3.1 requires 70 seconds for an 8-second clip, MiniMax H3 runs at 180 seconds, ByteDance's Seedance 2.5 at 317 seconds, and Kuaishou's Kling 3.0 Pro at 398 seconds2

2

. LTX claims roughly one-eighth the cost and one-seventh the render time of comparable models.

Understanding World Models and Their Applications

LTX-2.5 is a world model, fundamentally different from large language models. World models translate the physical world into a language that generative AI can understand, providing grounded comprehension that reduces hallucinations1

1

. This technology enables robots to navigate rooms without collisions and ensures AI-generated video clips show people walking around objects rather than fading through them. Companies can use world models to simulate digital environments, generate content, and program robots. The model includes a pretrained checkpoint specifically tuned for physical AI and robotics, giving teams a foundation to fine-tune on domain data that differs from cinematic video2

2

.

Technical Improvements Transform Generation Pipeline

LTX-2.5 rebuilds nearly every stage of the generation pipeline rather than adding capabilities to an older core. The model introduces a new diffusion video decoder that reduces visual artifacts in high-motion footage and reconstructs fine detail like text and faces while preserving high compression ratios2

2

. Native multishot generation renders full sequences as single outputs, maintaining character, scene, and voice consistency across cuts instead of stitching individually generated shots together. The model features a custom Gemma 4 language backbone and dedicated prompt enhancer for accurate handling of complex, multi-subject prompts. Through optimization efforts with Nvidia, a substantially improved distilled model delivers near-full-model quality at lower cost and faster inference, running locally on Nvidia RTX GPUs with reduced memory requirements. The pixel diffusion process contributes to better consistency across generations1

1

.

Quality Metrics Show Competitive Performance

In blind, side-by-side human preference tests where evaluators voted on videos from identical prompts without knowing which model produced them, LTX-2.5 recorded a 67% win rate2

2

. This narrowly surpassed Seedance 2.5 at 65%, with Gemini Omni Flash at 55%, MiniMax H3 at 50%, Seedance 2.0 at 44%, Wan 2.6 at 42%, and FLUX 3 at 28%. LTX labels these preference results as preliminary, noting expectations for evolution as evaluation expands. The company emphasizes these are vendor-reported measurements that buyers should test against their own workloads.

Deployment Flexibility Targets Enterprise Adoption

LTX-2.5 runs on any GPU with a minimum of 16GB of VRAM, enabling local inference on accessible hardware2

2

. Companies can deploy on-premises, at the edge, or via API, with no mandatory branding on output. The model supports fine-tuning on customer data and intellectual property. Movie studios could use LTX-2.5 to create and edit video clips featuring their characters while keeping proprietary assets secure1

1

. The model is free to use for organizations under $10 million in annual recurring revenue; larger companies negotiate licenses. Available now as open weights on Hugging Face, inside ComfyUI through a strategic day-one launch partnership, and through the LTX API for teams wanting managed generation2

2

.

Open-Weights Strategy Challenges Closed API Dominance

LTX models have surpassed 33 million downloads on Hugging Face, making the LTX family the most-used open world model line on the market2

2

. CEO and co-founder Zeev Farbman argues that neither government nor private companies make good stewards of private AI models, with one motivated by shifting political goals and the other by pure profit1

1

. "Decisions belong with the end users, through models they can control and inspect for themselves," Farbman writes. This philosophy aligns with Meta CEO Mark Zuckerberg's recent essay advocating that open-weight models will bring AI benefits to more people. Meta released open-weights model Muse Glimmer and promised to release weights of Spark 1.2. Popular Chinese models like Kimi 3 from Moonshot are being released in open formats, giving developers, researchers, and cybersecurity experts insights into functionality. While OpenAI and Google each have open-weight model families—GPT-OSS and Gemma 4 respectively—their closed models far outnumber open-weights ones. Anthropic has no open Claude models1

1

.