Open-source AI

Open-source AI is AI that its makers publish so that anyone may use it, look inside it, change it and pass it on, usually for free. A business can run such a model on its own computer or server, so customer data stays in-house and there is no fee per question. It does pay for the hardware, and the licence sets what it may do.

What counts as open source?

The term comes from software, where the source code is public under a licence that lets anyone use, study, change and share it. For AI, the Open Source Initiative (OSI), which maintains that definition for software, published the Open Source AI Definition 1.0 on 28 October 2024. It asks for the same four freedoms and for everything you need to change the system: the model's parameters (its weights), the complete code to train and run it, and enough information about the training data that a skilled person could build a substantially equivalent system.

Many models called open publish only the trained weights: you can download and run the model, but you cannot see how it was made. The OSI calls these open weights (see open weights vs open source). In everyday use, "open-source AI" covers both.

How big is the open ecosystem?

Hugging Face, a platform where developers share models and datasets, grew in 2025 to 13 million users, more than 2 million public models and more than 500,000 public datasets, according to its report State of Open Source on Hugging Face: Spring 2026 (March 2026). In the year before the report, Chinese models took 41% of downloads, more than any other country, after DeepSeek released its R1 model in January 2025 (see Why is China ahead in open-source AI?). Independent developers, who mostly compress, adapt and republish existing models, went from 17% of downloads before 2022 to 39% in 2025.

Can open models keep up with closed ones?

In April 2026 the Chinese lab Z.ai published GLM-5.1, a model of 754 billion parameters, with its weights on Hugging Face under the MIT licence. In Z.ai's own benchmark table it scored 58.4 on SWE-Bench Pro, a test with programming tasks from real software projects, just above GPT-5.4 (57.7) and Claude Opus 4.6 (57.3). On most other tests in the same table, a closed model from OpenAI, Anthropic or Google scored highest. The weights alone take 1.51 TB (Simon Willison). According to Hugging Face, small models are downloaded far more often than very large ones, because of cost, speed and available hardware.

How do you use an open model?

You download the weights, often from Hugging Face, and run them with software such as Ollama, llama.cpp or vLLM. Quantization stores the weights with fewer bits, so a model fits on smaller hardware. To teach a model your own material you fine-tune it, often with PEFT, which trains a small add-on instead of the whole model, using libraries such as LLaMA-Factory or Unsloth.

Example

Suppose a small software company wants a coding assistant that knows its own code, without sending that code to an outside service. It downloads DeepSeek-Coder 6.7B, trains an add-on on its own repositories with PEFT, and runs the model on its own server. The code stays in-house, and instead of API fees the company pays for the server and its upkeep. The model licence allows commercial use but lists uses that are forbidden. The OSI definition requires use for any purpose, so by that definition DeepSeek-Coder does not count as open source.