Open Source AI refers to AI systems — including models, training code, and sometimes training data — whose components are publicly released under licenses that permit inspection, modification, and redistribution. The term covers a spectrum: at the permissive end are releases like Meta's Llama 3 or Mistral's Mistral-7B, where model weights and inference code are freely downloadable; at the more restricted end are "source available" releases that impose commercial or derivative-work limitations. The distinction matters practically because it determines whether a team can fine-tune a model on proprietary data, run it on private infrastructure, or redistribute a customized version.

Think of it like open-source software in traditional programming. Just as Linux lets system administrators inspect and patch the kernel rather than waiting for a vendor, open-source AI lets researchers and engineers examine model weights, audit behavior, and adapt the model for specialized tasks without depending on a single provider's API.

The contrast with closed models is direct. GPT-4 is accessible only through an API — you send inputs, receive outputs, and never see the underlying weights. An open-weight model like Llama 3 70B can be downloaded, quantized to fit on a consumer GPU, and served locally. Teams with strict data-privacy requirements or unusual domain needs often fine-tune open models rather than expose sensitive data to third-party APIs:

# Pull and run a Llama 3 8B model locally with Ollama
ollama pull llama3
ollama run llama3 "Summarize the following contract clause: ..."

Open Source AI intersects closely with Fine-tuning, since publicly released weights are a prerequisite for most fine-tuning workflows, and with LoRA and Quantization, which make running large open models feasible on modest hardware. The Foundation Model concept is also central, as open releases are almost always large pretrained base models.

You will encounter the open-source AI distinction whenever evaluating build-vs-buy decisions, assessing data-privacy risk, or reading research papers that compare proprietary and open model performance on benchmarks.