# Open vs. Closed Models: What the Difference Means for You > An open model publishes its weights so you can download and run it; a closed one is used via API. What changes in control, cost, privacy and licenses. *Tags: ai, models, licenses, architecture* --- > [!info] Quick Definition > An **open model** (in practice, an open-weight model) publishes its trained parameters so anyone can download and run them. A **closed model** keeps its weights private and is used through an API or a product. The difference defines who controls the model, where it runs and how it is paid for. ## What does "open" mean in AI? The term is used less precisely than it seems. **Weights** are the parameters a model learned during training; with them and the right software, you can run the model. An open-weight model publishes only that. The Open Source Initiative (OSI) asks for more before calling something open-source AI. Its 1.0 definition requires four freedoms (to use, study, modify and share the system for any purpose) and three components: sufficiently detailed information about the training data so a skilled person can build a substantially equivalent system, the complete code for training and running the system, and the model parameters. So a model can be "open" in the sense of downloadable weights without meeting the full definition. Check what is actually published and under which license. | Type | What you get | Example | |---|---|---| | Closed | Access via API or product; no weights | Claude models, available through Anthropic's API | | Open weights, permissive license | Downloadable weights, commercial use without major strings | OpenAI's gpt-oss, under the Apache 2.0 license | | Open weights, custom license | Downloadable weights with conditions | Meta's Llama, under its community license | | Open source per the OSI | Weights, training code and sufficient information about the data | Each model has to be checked against definition 1.0 | ## What licenses will you run into? The license matters more than the "open" label. Two real cases: - **gpt-oss.** OpenAI released gpt-oss-120b and gpt-oss-20b under Apache 2.0, a permissive license. According to its Hugging Face model card, the 120b model runs on a single 80 GB GPU and the 20b model within 16 GB of memory. - **Llama 4.** Its community license requires you to prominently display "Built with Llama" on a related website if you distribute the materials or a product containing them, and to begin the name of any model you train or fine-tune with its materials or outputs with "Llama." If your products exceed 700 million monthly active users, you must request a license from Meta, which can refuse it. The OSI maintains that the Llama 3.x license is still not open source, because it restricts use for certain purposes and discriminates against some users. The practical lesson for you is to read the license before building on the model, rather than trusting the category name. ## What changes for a product? - **Control.** With downloaded weights you can pin the version and fine-tune the model. With a closed model you depend on the provider's lifecycle: Anthropic, for example, retires older models and commits to at least 60 days' notice for publicly released ones. That is a reasonable window, but it still forces you to migrate and re-test. - **Privacy.** An open model can run without sending data to third parties. Even so, security depends on your deployment, not only on the model. For the data picture, see [[privacy-in-ai-products|privacy in AI products]]. - **Cost.** An API charges per use (see [[cost-per-token|cost per token]]). An open model charges in infrastructure and people, or per token if you use it through an inference provider. - **Capability.** In May 2026, Epoch AI estimated that since January the most capable open-weight models had lagged closed frontier models by about four months on its Epoch Capabilities Index (ECI). That is an average on a composite index, not a guarantee for your task. Compare with your real cases, and remember that a smaller model may be enough (see [[small-vs-large-models|small vs. large models]]). - **Safety.** The NTIA's 2024 report notes that publishing weights broadens participation and reduces dependence on a few developers, but can also create risks to security, privacy and civil rights. And widely distributed weights cannot be withdrawn, and anyone can remove their safeguards. If your product relies on safety behavior, do not take it for granted. ## When should you NOT choose an open model? When you have nobody to operate it. Serving a model with good latency requires GPUs, monitoring, updates and testing. If that work is not part of your business, a closed API may be the sensible choice, even if it costs more per token. Also avoid it if you have not read the license: some impose attribution, naming or scale conditions, like Llama's. And if your volume is low, the fixed cost of a GPU rarely pays off. [[build-vs-buy-models|Build vs. buy]] works through that math. ## How do you decide? Ask yourself, in this order: 1. Can the data leave my infrastructure? 2. Do I need a fixed version of the behavior for years, not months? 3. Do I have a team that can operate the model? 4. What does it cost at my real volume? 5. Does the license allow my use? The answer usually points to one of three options: a closed API, an open model hosted by a provider, or an open model on your own infrastructure. The middle option is easier than it sounds: Hugging Face Inference Providers serves open models, such as gpt-oss-120b, behind an OpenAI-compatible API, pay-as-you-go and with no markup from Hugging Face on the provider's rate. To run one on your own machine, see [[local-ai|running AI locally]]. This is an architecture decision, but designers feel it: a model change can alter tone, format or interface speed. That is why design should take part in the evaluation, using a set of test cases like those in [[golden-datasets|golden datasets]]. ## Example Illustrative example: a clinic wants an assistant to summarize internal notes without them leaving its network. The team evaluates an open model on its own server. Summary quality is good for short notes, but long notes lose detail. The team decides to split notes into sections before summarizing and to keep the open model, because privacy is the main requirement. The interface shows each section separately so staff can verify the result. ## Common mistakes - Calling any downloadable model "open" without reading the license. - Choosing an open model without counting the cost of operating it. - Assuming local execution guarantees privacy without reviewing the rest of the system. - Switching models without re-evaluating the product. - Comparing an API's cost per token with GPU cost per hour without normalizing for volume. - Deciding from a general ranking instead of testing with your own tasks. ## Sources - [The Open Source AI Definition 1.0, Open Source Initiative](https://opensource.org/ai/open-source-ai-definition) - [Meta's LLaMa license is still not Open Source, Open Source Initiative (February 18, 2025)](https://opensource.org/?p=121585) - [Llama 4 Community License Agreement, Meta](https://dev.meta.ai/llama/llama4/license/) - [openai/gpt-oss-120b, model card on Hugging Face](https://huggingface.co/openai/gpt-oss-120b) - [Open models lag state-of-the-art closed models by 4 months, Epoch AI (May 29, 2026)](https://epoch.ai/data-insights/open-closed-eci-gap) - [Dual-Use Foundation Models with Widely Available Model Weights, NTIA report (July 30, 2024)](https://www.ntia.gov/issues/artificial-intelligence/open-model-weights-report) - [Model deprecations, Anthropic documentation](https://platform.claude.com/docs/en/about-claude/model-deprecations) - [Inference Providers, Hugging Face documentation](https://huggingface.co/docs/inference-providers/index) - [Open-weight model, Wikipedia](https://en.wikipedia.org/wiki/Open-weight_model) --- Source: https://www.fernandoux.com/en/wiki/ai/open-vs-closed-models/