Lightricks shipped LTX-2.3 on June 18, 2026, an open source 4K video generation model under the Apache 2.0 license. The release includes a 13 billion parameter diffusion transformer that produces up to 5 second clips at 24 frames per second. It supports text-to-video and image-to-video workflows. You can download the weights from Hugging Face and run the model locally. Unlike many free AI API tiers that changed pricing in 2026, LTX-2.3 has no per second fees or hidden credit limits. The full announcement is on the Lightricks homepage.
The model is the latest in a line of open source media releases from Lightricks. The company previously shipped HiDream O1, an open source image model. LTX-2.3 builds on that work but moves into high resolution video. The public release includes model weights, a reference Python implementation, a ComfyUI node, and LoRA fine-tuning code. Lightricks published the announcements on its own site and on Hugging Face. The code is available on GitHub. This is a full release, not a paper only.
Why does this matter? LTX-2.3 offers a real 4K open model at a time when many closed video APIs have added credits, watermarks, or usage caps. The Apache 2.0 license removes the vendor lock-in that comes with Sora, Veo, and Runway. You can run the model on your own GPU, keep your data local, and use the output commercially. The open release also puts price pressure on closed video tools. The wider shift is covered in free AI pricing changes June 2026.
The key technical facts are simple. LTX-2.3 uses a 13 billion parameter diffusion transformer. It generates 4K clips up to 5 seconds long at 24 frames per second. Lightricks reports an 84.2 on VBench-HD. The model runs on a single consumer GPU with 16GB VRAM for 1080p and 24GB for 4K. You can read more about the open source model landscape in best open source LLM models 2026. This release sits at the intersection of open weights, local control, and 4K video generation.
How Do the Top Options Compare?
| Model | Open License | Max Resolution | Max Clip Length | Hardware |
|---|---|---|---|---|
| LTX-2.3 | Apache 2.0 | 4K | 5 seconds | 16GB VRAM minimum |
| LTX-Video 2 | Apache 2.0 | 1080p | 8 seconds | 12GB VRAM |
| HiDream O1 | Apache 2.0 | 1440p | 5 seconds | 16GB VRAM |
| Closed video APIs | Proprietary | 4K | 10-60 seconds | Cloud only |
LTX-Video 2 figures are from Lightricks previous release. Closed API specs vary by provider and often change. Hardware values are for local generation, not cloud inference.
1. LTX-2.3 Model Overview , Best for local 4K video generation without API fees
Lightricks released LTX-2.3 on June 18, 2026 as an open source video generation model. The model uses a 13 billion parameter diffusion transformer architecture. It generates 4K resolution clips up to 5 seconds at 24 frames per second. The public release includes model weights, a LoRA adapter for fine-tuning, and a reference implementation. Lightricks published the release on its Lightricks homepage and on Hugging Face.
This is not a small research preview. LTX-2.3 is a full 4K video model with temporal attention layers that keep objects stable across frames. The model supports text-to-video and image-to-video. Lightricks says the model was trained on licensed and public domain footage. It avoids the hidden cost structure that many free AI video tools introduced in 2026. You can read more about the wider open source shift in state of open source on Hugging Face.
The model is designed for local use. You download the safetensors weights and run them on your own GPU. That means no data leaves your machine and no vendor can change the pricing later. The package also includes a Gradio demo script for a simple web interface. Lightricks does not host a public demo, so you need to install it yourself.
Key strengths:
- ✅ Generates true 4K video without a paid API
- ✅ Open weights allow local fine-tuning and commercial use
- ✅ Text-to-video and image-to-video modes included
- ✅ Runs on a single 16GB consumer GPU
- ❌ Maximum clip length is only 5 seconds
- ❌ Requires significant VRAM for 4K output
- ❌ First release lacks a hosted demo
Who it’s for: Developers and creators who need a free, local 4K video generation model without per-clip API fees.
2. LTX-2.3 Performance and Benchmarks , Best for measurable open source video quality
LTX-2.3 scores 84.2 on VBench-HD, the standard benchmark for high resolution video generation. Lightricks says this beats the previous LTX-Video 2 by 9 points. The model also scores 78.5 on EvalCrafter for text alignment. These numbers put it close to closed models like Runway Gen-4 and Google Veo 3 on quality, but with no usage limits. The full benchmark card is on the Lightricks website. The open source community can reproduce the scores because the evaluation code is public.
The model uses 13 billion parameters but samples with 25 diffusion steps. That keeps generation time at around 40 seconds per 5 second clip on an RTX 4090. At 4K resolution, generation can take 90 seconds. Lightricks includes a low VRAM mode that offloads the text encoder and VAE decoder. That allows 1080p output on 12GB GPUs. For comparison, many free AI API tiers now cap video seconds or add watermarks. The shift is covered in AI free tier limits get tougher.
Benchmark scores are useful but not the whole story. The model still struggles with long camera pans and complex physical interactions. You can get excellent results for single subject scenes, product shots, and stylized content. For multi character action scenes, closed models often remain stronger. The public benchmark card is a step toward honest comparison, but independent testing will take time.
Key strengths:
- ✅ Strong VBench-HD score of 84.2
- ✅ 25 step sampling keeps generation fast
- ✅ Low VRAM mode works on 12GB GPUs
- ✅ Evaluation code is public
- ❌ 4K generation is slow on midrange GPUs
- ❌ Benchmark scores are self-reported and need independent checks
- ❌ No built-in audio generation
Who it’s for: Researchers and creators who want a measurable open source video model with near-closed quality.
3. LTX-2.3 License and Openness , Best for commercial open source video without royalties
LTX-2.3 ships under the Apache 2.0 license. That means you can use the model commercially, modify it, and distribute your changes. You do not need to pay Lightricks a royalty or request a license key. The weights are downloadable from Hugging Face. The training code is not included, but inference and LoRA fine-tuning code is on GitHub. This is a major difference from closed video APIs that changed pricing in 2026. Many AI video tools introduced credits or tokens. LTX-2.3 removes that layer entirely.
One limitation is the training data license. Lightricks says the model was trained on licensed footage and public domain sources, but it has not released the full data manifest. That matters for companies that need to prove clean provenance. The model card does list what was excluded. You can compare this to other Apache 2.0 releases like Qwen 3.6. The key point is that the weights are open, not the training pipeline.
The open license has a real business impact. You can embed LTX-2.3 in a commercial product without negotiating a contract or sharing revenue. That is not true for Sora, Veo, or Runway. For startups and indie developers, this is the biggest reason to try LTX-2.3. The tradeoff is that Lightricks does not provide the same level of support or uptime as a cloud API.
Key strengths:
- ✅ Apache 2.0 allows commercial use and modification
- ✅ Weights are free to download from Hugging Face
- ✅ No API keys, credits, or per-second fees
- ✅ LoRA fine-tuning code included
- ❌ Training data manifest is not fully public
- ❌ Training code is not released
- ❌ Commercial use may still require your own IP review
Who it’s for: Startups and indie developers who need commercial friendly open video weights without vendor lock-in.
4. Running LTX-2.3 Locally , Best for local control and data privacy
To run LTX-2.3 locally, you need PyTorch 2.7 or higher and a GPU with at least 16GB VRAM for 1080p. For full 4K output, Lightricks recommends 24GB VRAM. The model uses ComfyUI and a standalone Python script. Installation is simple: download the safetensors weights from Hugging Face, install the diffusers version, and run a short command. You can also use the model in a Docker container to avoid dependency conflicts. This local setup means no data leaves your machine.
The reference script supports text-to-video, image-to-video, and keyframe conditioning. On a 24GB RTX 3090, a 1080p 5 second clip takes about 35 seconds. On a 4090, 4K takes around 90 seconds. Lightricks includes a Gradio demo script, but it does not host a public demo. That means you need to install it yourself. The local-first approach fits the broader open source AI trend described in running Llama 3 locally.
The model is not a one click installer. You need some comfort with Python and CUDA. But the documentation is clear and the community on Hugging Face is active. If you have used ComfyUI before, you will feel at home. If not, expect an hour or two of setup. For non technical users, a cloud API may be easier, but it will cost you.
Key strengths:
- ✅ Runs entirely on your own hardware
- ✅ ComfyUI and Python scripts included
- ✅ Low VRAM mode supports 12GB GPUs
- ✅ Docker image simplifies setup
- ❌ No hosted demo from Lightricks
- ❌ 4K requires high end consumer hardware
- ❌ First install can take over 30GB of disk space
Who it’s for: Technical creators and developers who want full control over video generation without cloud APIs.
5. LTX-2.3 Competitive Context and Next Steps , Best for open source advocates who need a 4K baseline
LTX-2.3 enters a crowded field. Closed models like Sora, Veo, and Runway still lead on long clips and physics accuracy. But those models charge per second and cap free users. LTX-2.3 gives you a fixed cost: your GPU. Other open source video models include Nvidia Cosmos 3 and HiDream O1. LTX-2.3 is stronger on 4K realism because of its temporal attention stack. It is weaker on very long video because of the 5 second cap.
The next steps matter. Lightricks says it will release a 10 second model and an audio generation module later in 2026. It also plans to add a LoRA marketplace. Those updates could close the gap with closed tools. But for now, the main tradeoff is clip length. If you need 30 second shots, you will need to stitch multiple generations. That is a real limitation, not a small one. The open release still matters because it pushes closed video APIs toward lower prices, a trend covered in AI price wars.
The open video space is also moving fast. Nvidia Cosmos 3 targets physical AI and robotics. HiDream O1 focuses on image generation. LTX-2.3 is the first major Apache 2.0 model to explicitly target 4K short form video. That is a specific niche, but it matters for ads, product demos, and social clips.
Key strengths:
- ✅ Open source pressure on closed video APIs
- ✅ Planned 10 second model and audio module
- ✅ Strong 4K realism compared to other open models
- ✅ Active community on Hugging Face and GitHub
- ❌ 5 second clip limit is a real workflow constraint
- ❌ Not yet competitive with Sora on long narration driven scenes
- ❌ Lightricks has not committed to a public benchmark leaderboard
Who it’s for: Open source advocates and video teams who can stitch short clips and want to avoid per-second API costs.
Frequently Asked Questions
Is LTX-2.3 free to use commercially?
Yes, the Apache 2.0 license allows commercial use, modification, and distribution. You do not need to pay Lightricks a royalty or request a license key. You still need to include the license notice and follow the terms.
What hardware do I need to run LTX-2.3?
You need a GPU with at least 16GB VRAM for 1080p generation and 24GB VRAM for full 4K output. A low VRAM mode supports 12GB GPUs for 1080p. CPU generation works but is extremely slow.
Where can I download LTX-2.3?
The weights are available on Hugging Face. Inference and LoRA fine-tuning code are on GitHub. The Lightricks homepage links to both locations. No API key or account is required.
How does LTX-2.3 compare to Sora or Veo?
LTX-2.3 offers open weights and no per clip fees. Closed models still lead on long clip consistency and physics accuracy. The main tradeoff is LTX-2.3’s 5 second maximum clip length.
Can I fine-tune LTX-2.3 on my own data?
Yes, LoRA fine-tuning code is included with the release. Full fine-tuning is not officially documented. Lightricks has not released the complete training data manifest.
Does LTX-2.3 generate audio?
No, LTX-2.3 generates video only. Lightricks says an audio generation module is planned for later in 2026. For now you need a separate audio model or tool.
What Should You Remember?
- Open source release: Lightricks shipped LTX-2.3 on June 18, 2026 under the Apache 2.0 license.
- 4K video: The model generates 5 second clips at 24 frames per second with a 13 billion parameter diffusion transformer.
- Local-first: It runs on 16GB VRAM for 1080p and 24GB for 4K, no API keys or credits needed.
- Benchmarks: Lightricks reports an 84.2 VBench-HD score, close to closed model quality.
- License advantage: Apache 2.0 means commercial use without royalties or per second fees.
- Current limits: The 5 second clip length and no audio generation are the biggest tradeoffs.
- What’s next: Lightricks plans a 10 second model and audio module later in 2026.
Free AI News is an independent editorial publication. Information about AI pricing, free-tier limits, and features changes frequently and may become outdated. Always verify current details through the vendor’s official pages. Affiliate links may earn a commission at no cost to you, and never affect our reporting.
