Open-source and open-weight models

In everyday use, almost any language model whose weights can be downloaded gets called open source. Weights are the numbers learned in training, the model itself. Whoever has them can run the model on their own hardware or with a hosting provider instead of calling a vendor's API. Strictly speaking, though, that only makes it an open-weight model.

The Open Source Initiative (OSI) set out what open source means for AI in October 2024 with the Open Source AI Definition 1.0. Under the new definition, an open AI system must grant four freedoms: use it for any purpose, study it, modify it and share it. That requires the weights plus the complete code for training and running the model, and information about the training data detailed enough for a skilled person to build a substantially equivalent system. Many well-known open models fall short of that.

For companies, the distinction is more than semantics. The license decides whether a model may be used commercially, modified and built into your own products. Access to the weights decides whether data has to leave the company and how tightly you are tied to a vendor. And whoever runs the model carries the cost and the work.

Open source, open weights and proprietary.

Language models come in three levels of openness. Proprietary models can only be used through the vendor's API or application. The weights stay with the vendor, as with the models behind ChatGPT, Claude or Gemini.

Open-weight models offer their weights for download. Training data and training code, however, mostly stay private. Most models considered open fall into this group, for example the Llama, Gemma, Mistral, Qwen and DeepSeek families and OpenAI's gpt-oss (as of November 2026). What you may do with the weights is set by the license, which ranges from very permissive to heavily restricted.

Open source in the OSI sense demands more. One example is OLMo from the Allen Institute for AI: its first release in 2024 included weights, training code, training data and intermediate training checkpoints. The OSI definition does not always require the data itself, but it does require a description that makes the training reproducible.

The label on a model page is therefore no proof. What counts is what was actually released and what the license allows.

The license decides how you can use it.

The Apache License 2.0 allows use, modification and redistribution, including commercially, and includes an explicit patent license. Obligations are modest: include the license and notices, and mark changes. OpenAI released gpt-oss under this license in 2025. The very short MIT License is similarly permissive.

Vendor licenses set their own terms. The Llama 4 Community License of April 2025 allows commercial use but requires a separate license from Meta if the licensee's products had more than 700 million monthly active users when the version was released. Anyone redistributing Llama must display "Built with Llama", and derived models must have names starting with "Llama". The license incorporates an acceptable use policy listing prohibited uses. For Llama 4's multimodal models, that policy withholds the license rights from companies with their principal place of business in the EU; end users of products that include such models are exempt. The OSI stated back in 2023 that the Llama license is not open source because it restricts user groups and fields of use.

Licenses change between versions, so check the text of each version before use.

What the EU AI Act provides for open models.

The EU AI Act eases some rules for free and open-source licenses, but narrowly. Under Article 2(12), the regulation does not apply to AI systems released under such licenses. That exemption does not cover high-risk systems, prohibited practices under Article 5 or systems with transparency obligations under Article 50.

For providers of general-purpose AI models, Article 53(2) waives two obligations: technical documentation for the authorities and information for downstream providers. The condition is a license that allows access, use, modification and distribution, plus public release of the parameters, including weights, and of information on architecture and usage. The copyright policy and the summary of training content remain mandatory, and the exemption does not apply to models with systemic risk. Under Recital 103, a model loses the exemption if it is provided for a price or otherwise monetized.

For companies that deploy an open model, little changes: they keep their deployer obligations, for example on AI literacy, transparency and, where relevant, high-risk use. The relief is for model makers, not users.

How an open model runs in practice.

An open-weight model runs once its weights are loaded into a runtime. The files usually sit on platforms such as Hugging Face. On a single machine, tools like Ollama or llama.cpp are common; for servers handling many parallel requests, inference servers such as vLLM.

Memory needs depend on parameter count and quantization, the precision at which the weights are stored. In mixture-of-experts models, only part of the network computes each token, but the whole model still has to sit in memory. According to its model card, gpt-oss-20b has 21 billion parameters, 3.6 billion of them active per token, and runs within 16 GB of memory.

There are three ways to operate. Your own hardware gives the most control over data and versions but needs your own staff for operations and security. A hosting or cloud provider runs the open model for you; then the usual service-provider questions apply: where data is processed and stored, and what the GDPR data processing agreement says. The third option is a proprietary vendor's API. In the first two, only the weights are open, not the operations.

What it costs, and what drives the cost.

An open license saves the license fee, not the operating costs. Self-hosting means paying for GPUs or GPU rental, power, setup, integration, security updates, monitoring and the people who run it all, plus testing and rolling out new model versions. These costs are largely fixed, however many requests the model answers.

An API, by contrast, bills per token and costs little at low usage. Running your own model therefore tends to pay off with high, steady volume, with data that should not leave the company, or when a model version has to stay unchanged for years. With low or highly variable load, an API is usually cheaper. A hosting provider offering open models sits in between.

A sound basis is the cost per completed task for each of the three options, calculated with your real volume and the quality the task demands. That requires logging the cost of every call.

Common misconceptions.

Open means free: the weights cost nothing, running them does.

Open means privacy-compliant: running a model yourself can improve data control, but it replaces neither a legal basis nor deletion rules, access rights or logging. Run by a hosting provider, an open model raises the same data protection questions as any other service.

Open means worse: whether an open or a proprietary model handles a task better depends on the task, the language and the model version. Only testing on your own cases, in your own languages, is reliable.

Open means transparent: having the weights does not tell you why the model answers the way it does. Without information on the training data, what it learned and which biases it carries stay unknown.

Download once and you're done: open models also need updates, protection against attacks such as prompt injection and review of their output. Whoever runs an open model takes on the work the vendor would otherwise do.

Where the term stops working

Open source and open weight describe the terms under which a model is shared, not its quality, its security or whether a particular use is lawful. They differ from open-source tools around models, such as runtimes, which are not models themselves, and from open datasets. Self-hosting is a separate matter too: some proprietary models are licensed for on-premises use, and open models often run with a hosting provider.