Short answer: yes, for most people. You can download the weights, run them locally, and use the output commercially without paying Lightricks anything.
Longer answer: "free" and "open source" are not the same thing, and LTX 2.3 sits precisely in the gap between them. The weights are open. The license is not a standard permissive one — it contains a revenue threshold that turns free into paid at a specific company size. If you are a solo creator, that threshold will never touch you. If you are deploying inside a company, someone needs to read it before you ship.
This guide covers what the license actually says, what running LTX 2.3 costs in practice, and which "free" route makes sense for which situation. Sourced from the LTX-2 Community License Agreement, the LTX-2.3 model card, and the LTX-2 repository, checked July 2026.
What the license actually says
LTX 2.3 is released under the LTX-2 Community License Agreement — Lightricks' own license, not Apache or MIT. Three provisions matter:
1. The $10M revenue threshold. The agreement states that entities with annual revenues of at least $10,000,000 are "Commercial Entities" and are required to obtain a paid commercial use license in order to use LTX-2 and its derivatives. Such entities are directed to contact Lightricks, and commercial use outside that arrangement is described as strictly prohibited.
2. Derivatives inherit the agreement. Any derivative — a fine-tune, a merged checkpoint, a distilled variant — must be distributed exclusively under the same agreement, and you must give notice to downstream users.
3. Attribution must be preserved. Attribution notices pertaining to LTX-2 have to be retained when you redistribute.
What the license does not do is stop an individual or a small company from using outputs commercially. If you are a freelancer making client videos, a YouTuber, a small agency, or a startup well under $10M in revenue, the practical answer is: use it, keep attribution intact if you redistribute weights, move on.
This is a genuinely different posture from, say, Wan 2.2's Apache 2.0. Neither is better in the abstract — but if your organization's policy is "permissive licenses only," LTX 2.3 requires an actual decision rather than a rubber stamp.
The obvious caveat: this is a summary written by someone reading the text, not legal advice. The full agreement is roughly 21 KB in the repository. If your company is anywhere near the threshold, have counsel read the original.
The dependency people forget: Gemma
LTX 2.3 does not ship its own text encoder — it uses Gemma 3 12B, pulled from Google's repository. That model carries Google's own terms of use, separate from Lightricks' license.
So a compliant local deployment is governed by two agreements, not one. This almost never matters for individuals and occasionally matters a lot for enterprises. It is worth knowing before someone in legal discovers it after launch.
What "free" actually costs
No license fee is not the same as no cost. The real bill:
| Cost | Reality |
|---|---|
| Download | 46.1 GB per checkpoint. The full repository is ~157 GB. |
| Disk | Minimum viable distilled setup is ~47 GB plus the text encoder; a full dev + distilled + upscaler + LoRA setup roughly doubles that. |
| GPU | The bf16 checkpoints are 46.1 GB. FP8 quantization and CPU/disk offload exist precisely because most cards cannot hold that. |
| Time | Offloading to CPU or disk works — the docs explicitly note disk streaming is slower. You are paying in wall-clock. |
| Electricity | Local generation is not free; it is just billed by your utility instead of an API. |
| Training | LoRA training wants 80 GB+ VRAM in standard configs, or 32 GB with the INT8 low-VRAM config. That is usually a rental. |
The honest framing: LTX 2.3 is free of license fees and expensive in hardware. For someone who already owns a capable GPU, that is a fantastic deal. For someone who does not, "free" can quietly mean a four-figure purchase.
The four ways to use LTX 2.3 without paying Lightricks
1. Local via the Python package. Clone LTX-2, download the distilled checkpoint plus the spatial upscaler plus Gemma, run ltx_pipelines.distilled. Maximum control, maximum setup.
2. Local via ComfyUI. Six native LTX-2.3 templates, no custom nodes required, FP8 checkpoint builds that are lighter than the bf16 files. The most practical local route for most people.
3. LTX Desktop. Lightricks ships an open-source desktop app for generating video with LTX models locally — the friendliest local option if you would rather not touch a node graph or a terminal.
4. Browser-based generation. No download, no GPU, no CUDA version to chase.
Route 4, concretely: ltx23.app runs LTX 2.3 in the browser. It is the sensible first stop — check whether the model does what your project needs before you commit 47 GB of disk and an evening of setup to finding out.
Which route is right depends less on budget than on what you are optimizing for: control (1), balance (2), simplicity (3), or zero setup (4).
"Free" vs "open source": why the distinction matters
People search is ltx 2.3 open source and is ltx 2.3 free interchangeably, but they are different questions with different answers.
- Open weights? Yes — downloadable, inspectable, runnable, fine-tunable.
- Open source by OSI definition? No. The revenue threshold and the copyleft-style derivative clause are not part of a standard OSI-approved license.
- Free to use? Yes, below $10M annual revenue.
- Free to build a business on? Yes, until your business crosses that line — then it is a conversation with Lightricks.
That structure is increasingly common in AI model releases: permissive for the community that builds ecosystem value, commercial for the enterprises that can pay. Worth understanding on its own terms rather than flattening it into "open source" or "not open source."
FAQ
Is LTX 2.3 free to download? Yes. The Hugging Face repository is gated — accept the terms and use a read token with gated-repo access — but there is no charge.
Can I sell videos made with LTX 2.3? Nothing in the license restricts commercial use of outputs by entities below the revenue threshold. Above $10M in annual revenue, a paid commercial license is required to use the model.
Is LTX 2.3 open source? Open weights under a custom community license, not an OSI-approved open-source license.
Do I need to credit Lightricks in my videos? The attribution requirements in the license concern redistribution of the model and its derivatives. Output attribution is a separate question — check the agreement for your specific case.
Is there a free hosted option? Yes, browser-based generation exists, including at ltx23.app. Lightricks also runs an official playground.
What if I fine-tune it? Derivatives must be distributed under the same agreement, with notice to downstream users.
Does the text encoder have its own license? Yes — Gemma 3 12B comes from Google with Google's own terms.
Bottom line
LTX 2.3 is free in the way that matters to almost everyone reading this: download it, run it, sell what you make. The one thing to actually check is the $10,000,000 revenue threshold, and the one cost people underestimate is hardware, not licensing.
If you are not sure the model is worth 47 GB of your disk yet, answer that question first at ltx23.app — then download deliberately.
Sources
Verified July 2026 against primary documentation:
- LTX-2 Community License Agreement (repository copy) — $10,000,000 revenue threshold, derivative distribution, attribution requirements
- Lightricks/LTX-2.3 — Hugging Face model card — license name, file sizes, gated access
- Lightricks/LTX-2 — official GitHub repository — download instructions, token requirements, official playground link
- LTX-Desktop — Lightricks' open-source desktop app — local generation without a node graph
- ComfyUI official LTX-2.3 tutorial — native templates, FP8 builds
- google/gemma-3-12b-it-qat-q4_0-unquantized — text encoder and its separate terms
- ltx-trainer documentation — 80 GB+ / 32 GB INT8 training requirements
- LTX-2 optimization documentation — FP8 and offload as the cost-reduction path
- LTX-2 installation & CLI flags —
--offloaddisk/CPU trade-offs - ltx.io — official LTX site — product and API entry point
- LTX-2 technical paper (arXiv 2601.03233)
This article summarizes license text for orientation only and is not legal advice. Read the agreement in the repository before making commercial deployment decisions.
