LTX, the Jerusalem-based generative AI company spun out of Lightricks, is releasing LTX-2.5, the newest version of its open-weights world model. The company is positioning the release less as a video-generation upgrade and more as an infrastructure decision: run a frontier-grade model on your own hardware, keep your data and IP in-house, and skip the per-generation billing that comes with closed, hosted alternatives.
That framing matters more than the spec sheet. AI video tools have mostly competed on output quality. LTX is trying to compete on something enterprises actually lose sleep over: who controls the model, where the data lives, and what happens to their IP once it touches someone else's cloud.
What's new in 2.5
LTX-2.5 is built on the same open-weights foundation as its predecessors, but the company says it rebuilt most of the generation pipeline rather than layering new features onto the old core. The headline changes:
- A new diffusion video decoder aimed at reducing visual artifacts during high-motion scenes, while preserving LTX's compression advantage and staying closer to the source footage.
- Native multishot generation, which renders an entire sequence as a single output instead of stitching shots together — the goal being consistent characters, scenes, and voice across cuts.
- A custom Gemma 4 language backbone paired with a dedicated prompt enhancer, meant to improve how the model interprets complex, multi-subject prompts.
- A pretrained checkpoint for physical AI and robotics, giving teams a base they can fine-tune for environments that look nothing like cinematic video — warehouse floors, offshore rigs, factory lines.
- A more efficient distilled model, which LTX says delivers close to full-model quality at lower cost and faster inference, built together with NVIDIA for local inference on RTX GPUs.
LTX claims the result runs at roughly one-eighth the cost and a fraction of the render time of comparable closed models. On the speed side, LTX put a number behind that: 6.8 seconds to generate a 10-second, 720p clip on-prem with two GB200s. For comparison, the company's own API tier comes in at 23.7 seconds, Gemini Omni Flash at 52 seconds, and Kling 3.0 Pro at nearly 400 seconds for the same task.
These are LTX's own internal benchmarks, measured against third-party providers via fal.run, not independently verified numbers. Worth pressing LTX on methodology if you're evaluating this for a production workload, but even with that caveat, the gap between on-prem inference and the hosted competition is wide enough to be the more interesting story than the raw multiplier LTX is advertising.
The open-weights argument
LTX's core pitch hasn't changed since its 2024 spinout from Lightricks: ship the model as open weights, let enterprises fine-tune it on their own assets, and let them run it locally or in a private cloud. LTX-2.5 is available on Hugging Face, natively inside ComfyUI as a launch partner, and through the LTX API for teams that want a managed version instead. It's free for organizations under $10 million in annual recurring revenue — a threshold that puts real pressure on how LTX plans to monetize the model once a customer grows past it.
CEO Zeev Farbman frames the openness as a response to a problem world models face that large language models mostly didn't: holding motion, space, and sound consistent over time is a harder problem than predicting the next word, and he argues that keeping the model open is what lets teams own their hardware and their output rather than renting access to someone else's.
Where that argument gets concrete is in how LTX-2.5 stacks up against the alternatives on the terms that actually govern enterprise deployment — not benchmark scores, but jurisdiction, licensing, and what you're allowed to do with the weights once you have them.
A few rows worth flagging. LTX and Veo/Grok are the only options governed under U.S. law; MiniMax H3, Seedance, and Kling fall under Chinese jurisdiction — a real consideration for any enterprise with data residency or compliance requirements. MiniMax H3's license requires mandatory branding on outputs, something LTX doesn't impose. And LTX is the only model in the group that clears every row: open weights, no mandatory branding, fine-tuning rights on your own data, and a minimum VRAM low enough (16 GB) to run on a single consumer-grade GPU rather than a multi-GPU cluster.
Again, this is LTX's own comparison, built from public model cards and licensing terms — not a neutral third-party audit. But unlike the speed and cost numbers, most of what's in this table (governing jurisdiction, branding requirements, license terms) is independently checkable against each vendor's own published documentation, which makes it a fairer table to lean on.
Who's already building on it
LTX says the launch comes with three partnerships aimed at three different use cases. Asteria is using LTX for original film and video production. Reactor, a developer platform for real-time generative video, is running LTX-2.5 on its low-latency inference stack to power interactive avatars and live, responsive worlds — the kind of workload that can't tolerate round-trips to a hosted API. And ComfyUI, the node-based environment much of the open-source AI art community already builds in, is a day-one launch partner, meaning LTX-2.5 is available inside that workflow from the moment the embargo lifts.
LTX says it has more than 33 million downloads across all model versions on Hugging Face, which it points to as evidence it's the most-used open world model on the market. That's a real number worth noting, but download counts don't tell you how many of those downloads turned into production deployments — a distinction worth pressing LTX on directly.
The questions worth asking next
The open-weights pitch is compelling on paper. Whether it holds up depends on details LTX hasn't fully spelled out yet: what "production-ready" means in terms of uptime, throughput, and support; which parts of the stack customers can genuinely self-host versus what stays tied to LTX's hosted services; how rights, provenance, and indemnification work for commercially generated video; and what happens to pricing once a company crosses that $10 million revenue line.
LTX has spent two years building the case that open weights are the more trustworthy way to bring generative video into a business. LTX-2.5 is the company's most direct attempt yet to prove that case with production partners instead of just promises. The next test is whether the enterprises watching from the sidelines are convinced enough to build on it themselves.
