What are open weights?
When OpenAI releases GPT-4, you get API access. When Meta releases LLaMA, you get the weights. That distinction determines what you can do with the model, where you can run it, and what it costs per generation.
Definition
Open weights refers to the public release of a trained neural network's parameters. When weights are open, anyone can download them, run inference locally, fine-tune the model on their own data, and integrate it into applications without going through a paid API.
This is distinct from open source in the traditional software sense, which typically means releasing source code. Many open-weight models do not release training code or data, only the trained parameters. The term "open weights" is increasingly preferred over "open source" for AI models to reflect this distinction accurately.
What open weights enables
Running a model through a commercial API means paying per request, sending your data to a third-party server, and operating within the provider's rate limits and availability guarantees.
Open weights removes these constraints. With access to the weights, you can run inference locally at no per-generation cost beyond electricity and hardware, keep your data and generated content on your own infrastructure, fine-tune on proprietary data without it leaving your environment, and integrate the model into production systems without vendor lock-in. You can also quantize or optimize the model for your specific hardware.
For enterprise users with IP sensitivity, local inference on open weights means generated content, prompts, and source assets never touch a third-party server. This is the difference between a model a procurement team can approve and one they cannot.
Open weights vs. closed models
Closed models release only API access. The weights stay proprietary. Users interact through standardized endpoints, cannot customize the architecture, and depend on the provider for availability and pricing.
Open-weight models exist on a spectrum. Some release weights with permissive licensing allowing commercial use, modification, and redistribution. Others restrict commercial use or redistribution. Licensing terms matter as much as the weights themselves for production deployment decisions.
A brief history
The large language model community shifted significantly toward open weights starting in 2023. Meta's LLaMA release (February 2023) catalyzed the open-source AI community, enabling fine-tuning and local deployment at a scale not previously possible. Subsequent releases from Meta, Mistral, and others continued expanding the ecosystem.
For image generation, Stable Diffusion (August 2022) was the first major open-weight model at scale. It triggered a community ecosystem of tools, custom models, and fine-tuned variants that reshaped the field. For video generation, LTX-2's open-weight release in late 2025 and early 2026 achieved approaching 5 million downloads on Hugging Face, believed to be the fastest adoption ever recorded on the platform.
Why open weights matter for video generation
Video generation models are large. Fine-tuning them for brand-specific style or character consistency through an API is often impractical; providers rarely expose training infrastructure. Open weights make fine-tuning possible.
They also enable community development: custom integrations, quantization for lower-hardware deployment, ComfyUI nodes, custom sampling schedulers, and research extensions the original developers did not anticipate.
LTX-2 open weights
The LTX-2 open-weight model is available on Hugging Face under a permissive license. It runs locally via LTX Desktop on consumer-grade GPUs at zero per-generation cost, or integrates into custom pipelines through Diffusers, ComfyUI, or direct PyTorch.
