News · 2026-08-11

LTX-2.5 ships open weights and a chart that races its own hardware

Lightricks released LTX-2.5 on 11 August 2026 with open weights on Hugging Face, free commercial use for anyone under $10M in annual revenue, and a rebuilt generation stack. The headline change is architectural rather than incremental: a new diffusion video decoder replaces the older reconstruction stage, and it allocates rendering compute according to how complex each part of the scene is. The company's own comparison chart also claims a 10-second clip in 6.8 seconds, with a methodology footnote that deserves as much attention as the number.

Key facts

• Released 11 August 2026, with a same-day API changelog enabling text-to-video, image-to-video and audio-to-video endpoints.
• Free commercial and production use under $10M annual revenue, no mandatory branding, with fine-tuning rights.
• Minimum 16GB VRAM to run locally, per Lightricks' own comparison table -- unchanged from the previous generation.
• Primary sources: the LTX-2.5 model page, the open weights on Hugging Face, and the API changelog for 11 August.

What actually changed

Lightricks lists four changes from LTX-2.3, and together they amount to a rewrite of everything around the model rather than a bigger model. A new diffusion video decoder replaces the previous reconstruction stage. A Gemma 4 12B text encoder with a custom prompt enhancer now handles the instruction. An automatic duration predictor chooses how long the clip should be based on the action requested. And the distilled fast model was retrained on a larger dataset with reinforcement learning.

The decoder change carries the marketing name "Diffusion Fidelity Rendering," which the model page describes as "a new video generation technology that allocates rendering compute by scene complexity." Strip the branding and the idea is sound and familiar from other parts of the field: not every frame deserves equal effort. A static wide shot of a wall needs less sampling work than a face turning through changing light. Spending a fixed budget uniformly across a clip wastes compute on the easy parts and starves the hard ones. See diffusion models and diffusion distillation for the underlying machinery.

The other headline features -- native multishot generation that holds character, environment, lighting and voice across connected shots, native 4K HDR, and a RAW workflow -- point at a specific ambition. Lightricks is not positioning this as a clip toy. It is positioning it as something that survives contact with a professional colour and finishing pipeline.

The chart, and the note underneath it

The comparison table is going to be the most-screenshotted artefact of the release. It lists a 10-second image-to-video clip at 6.8 seconds for LTX 2.5 on-prem, 23.7 seconds through the LTX API, and then a long tail: 70 seconds for Veo 3.1, 180 seconds for MiniMax H3, 398 seconds for Kling 3.0 Pro.

Directly beneath it, Lightricks explains how those numbers were produced: "On-prem runs on our own hardware, 2x GB200, at steady state. API figures are measured end to end on fal.run, a third party provider, and include queue time... Resolutions vary."

That is a steady-state run on two of the most expensive accelerators money can buy, timed against competitors' queue-inclusive third-party API latency at different resolutions. It is not a controlled benchmark and it does not claim to be. The credit due here is real -- the disclosure is on the same page, in plain language, which is more than most vendors manage. The criticism is equally real: the chart is engineered to be lifted without its footnote, and it will be.

The number local users care about

The same table lists 16GB minimum VRAM. Anyone hoping 2.5 dropped into a lower hardware class should recalibrate. What improved is packaging and configurability -- the release ships as a split pack with separate transformer, text encoder, video and audio components rather than one monolith -- not the floor. The distilled path runs a fixed 8-step schedule at guidance 1, which is why it is the speed path rather than simply a smaller copy of the full model. Background: classifier-free guidance.

Why it matters

The licence is the story for builders. Free commercial and production use under $10M in revenue, no mandatory watermark branding, and the right to fine-tune on your own data puts LTX-2.5 in a category almost nothing else in video occupies. API pricing is published and per-second -- $0.09 at 720p rising to $0.37 at 4K -- which means the self-host versus API decision is now an ordinary economic calculation rather than a forced choice. That matters more than any benchmark: the open-weight lane in video has spent two years being a compromise. It is starting to be a preference.

The honest caveat

Nothing in the release promises that fine-tuned adapters from LTX-2.3 carry over. Given a new decoder, a new text encoder and a new video autoencoder, the safe assumption is that they do not until an adapter author says otherwise. And Lightricks' documentation is not internally synchronised: the model and pricing pages describe retake and extend for 2.5 while the support matrix still lists only the three basic endpoints. That reads as staged docs rather than a coherent release snapshot -- worth checking against the API before you build on a feature you read about on a landing page.

Primary source, verified: read the paper →

Ground Truth. (2026, August 11). LTX-2.5 ships open weights and a chart that races its own hardware. Ground Truth. https://groundtruth.day/news/ltx-2-5-ships-open-weights-and-a-chart-that-races-its-own-hardware.html

@misc{groundtruth:ltx-2-5-ships-open-weights-and-a-chart-that-races-its-own-hardware,
  title  = {LTX-2.5 ships open weights and a chart that races its own hardware},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {aug},
  url    = {https://groundtruth.day/news/ltx-2-5-ships-open-weights-and-a-chart-that-races-its-own-hardware.html}
}

Topics: video-generation · open-weights · diffusion · lightricks · model-release