Open or proprietary AI models: how to choose

Anyone starting to use artificial intelligence in a project soon runs into a choice: use a model offered as a service by a company, accessed over the internet, or download a model and run it on your own servers.

The debate is usually framed as “open source versus proprietary”. The label is not much help. There are models with very different degrees of openness, and licenses with very different rules. This article sorts out those differences and proposes criteria for deciding. If you want to review what a language model is, see what LLMs are.

Three degrees of openness

A language model is, in the end, a file with billions of numbers, called parameters or weights, adjusted during training. To understand openness, it helps to separate three things: the weights, the code used to train and run the model, and the training data.

With that separation, three situations can be distinguished:

• Proprietary (closed) model: you do not receive the weights. You use the model through an app or an API, which is the interface through which a program sends requests to the company’s server and receives responses.
• Open-weight model: the weights can be downloaded and run by you, under a license. The training data and part of the code are generally not published.
• Broadly open model: in addition to the weights, the training code and detailed information about what was used for training, or the data itself, are published.

What the definition of “open source” requires

The Open Source Initiative (OSI), the organization that maintains the definition of open source software, published the Open Source AI Definition, version 1.0. According to the text, an open source AI system must guarantee four freedoms: to use it for any purpose without asking permission, to study how it works, to modify it and to share it.

The definition says that exercising those freedoms requires access to the “preferred form” for making modifications. This includes three elements:

• sufficiently detailed information about the training data for a skilled person to build a substantially equivalent system;
• the complete code used to train and run the system;
• the model parameters.

By this criterion, many models promoted as “open source” are, more precisely, open-weight. In this article, we use “open weights” when what is available are the parameters, and “broadly open” when code and data are also available.

Real license examples

Licenses are what define, in practice, what you can do. Some examples, according to the official documents consulted on October 8, 2026:

• gpt-oss (OpenAI). The official repository includes two models, gpt-oss-120b and gpt-oss-20b, under the Apache 2.0 license, considered permissive. According to OpenAI, the larger one fits on a single 80 GB GPU, and the smaller one runs with 16 GB of memory.
• DeepSeek-R1. The model card states an MIT license for the code and the weights, with commercial use allowed, including for distillation (using the model’s outputs to train another one).
• Mistral Small 3. In its January 30, 2025, announcement, Mistral AI stated an Apache 2.0 license and said that, quantized (using lower-precision numbers to take up less memory), the model runs on a single RTX 4090 card or on a Macbook with 32 GB of RAM.
• Llama 3.1 (Meta). The Llama 3.1 Community License, dated July 23, 2024, allows commercial use, but with a condition: if, on the version’s release date, the licensee’s products had more than 700 million monthly active users, it must request a license from Meta, which may or may not grant it at its discretion. It is a license with restrictions, unlike the permissive ones.
• Olmo 3 (Ai2). The Allen Institute for AI publishes Olmo 3 under Apache 2.0 and, according to the model card, releases code, checkpoints (intermediate versions from training) and training details, and points to the Dolma 3 dataset. It is an example of broad openness.

Licenses change between versions. Before using any model, read the license for the exact version you are going to download.

Performance: the gap has narrowed

For a long time, closed models had a clear quality advantage. The AI Index 2025 report, from Stanford University’s Institute for Human-Centered AI, measured this gap on Chatbot Arena, a leaderboard based on comparisons made by users. In early January 2024, the best closed model outperformed the best open-weight model by 8.04%. By February 2025, the gap had fallen to 1.70%.

Two caveats. First, these figures are from February 2025, and the leaderboard changes every month. Second, a general ranking does not tell you how the model will perform on your task. For that, you need to test with real examples, as we explain in how to evaluate AI responses.

Criteria for choosing

In the newsroom’s assessment, the decision becomes clearer when you answer six questions.

1. Where does your data go?

With a proprietary model, the text you send travels to the company’s servers. The terms of use say what it can do with it. Anthropic’s commercial terms, in effect since June 17, 2025, for example, state that the company may not train models on content submitted by customers of those services. Consumer terms may differ, and each provider has its own.

With open weights running on your own infrastructure, the data does not need to leave it. For sectors with strict confidentiality rules, this may weigh more than any difference in quality.

2. Who will operate the model?

Running a model requires GPU servers, updates, monitoring and security. With an API, that falls to the provider. With open weights, it falls to you, or to a cloud provider you hire to host the model.

3. What is the total cost?

With an API, you pay per use, usually per token (the units of text the model processes; see tokens in AI models). With open weights, the model may be free, but the hardware, energy and engineering hours are not. In the newsroom’s assessment, low or variable volume tends to favor the API; high, stable volume may justify your own infrastructure. Do the math with your own numbers, as we suggest in AI ROI in business.

4. Do you need to adapt the model?

With the weights in hand, you can do fine-tuning, that is, continue training with your own examples to specialize the model. You can also run it offline or on local equipment. Some closed services offer fine-tuning, but within limits set by the provider.

5. What happens if the provider changes?

A closed service can change prices, retire versions or change rules. With open weights, the version you downloaded stays with you. On the other hand, security fixes and improvements then depend on your team.

6. What does the license allow?

Check whether the license allows commercial use, whether it imposes a user limit, whether it requires attribution and whether there is an acceptable use policy. The Llama 3.1 case shows that “can download” is not the same as “can use in any way”.

Risks and the safety debate

In the newsroom’s assessment, distributing open weights has a consequence that cannot be undone: once the files are in circulation, the author cannot recall them or control how they will be modified.

The paper On the Societal Impact of Open Foundation Models, by Sayash Kapoor, Rishi Bommasani and co-authors (2024), lists benefits of open models in innovation, competition, distribution of decision-making power and transparency. The same paper points to properties, such as greater customizability and limited monitoring of use, that generate both benefits and risks. The authors propose assessing marginal risk, that is, how much risk the open model adds compared with what already exists, such as search engines and available tools. For misuse such as cyberattacks and biological weapons, they conclude that the available research is still insufficient to measure that risk.

The US government reached a similar position. The NTIA report, dated July 30, 2024, concluded that the evidence at the time was not enough to justify restrictions on open-weight models and recommended actively monitoring the risks, without ruling out future measures.

For users, in the newsroom’s assessment, the risks are more concrete. A model downloaded from an unofficial source may have been tampered with. A model that is not updated may retain known flaws. And any model, open or closed, can be confidently wrong, as we show in AI hallucinations.

What about regulation?

In the European Union, Article 53 of the AI Act deals with providers of general-purpose AI models. Models released under a free and open license, with weights, architecture and usage information publicly available, are exempt from two technical documentation obligations. The copyright compliance policy and the public summary of the content used for training still apply. The exemption does not apply to models deemed to pose systemic risk.

In other words: even where there is a specific rule, “open” does not mean “no obligations”. And whoever uses the model in a product remains responsible for what the product does.

Key takeaways

“Open source” in AI is a loosely used label. What matters is what was published (weights, code, data) and what the license allows. Under the OSI definition, an open source model must release code, parameters and detailed information about the data.

Closed models deliver convenience, outsourced operation and, often, the best performance of the moment. Open-weight models deliver control over data, the ability to fine-tune and independence from the provider, in exchange for more technical work. The quality gap between the two groups has narrowed, according to the AI Index 2025.

In the newsroom’s assessment, the best path is to test two or three candidates on your real tasks, read each one’s license and calculate the total cost before deciding.