Open-Source AI

Artificial intelligence models, datasets, and training infrastructure released under licenses that permit public access, modification, and redistribution, as contrasted with closed, proprietary AI systems where weights, architectures, and training data are kept confidential. The definition of 'open-source' in AI is contested, as releasing model weights (allowing inference and fine-tuning) does not necessarily include training code, data, or documentation, and some releases include usage restrictions that deviate from traditional open-source principles. Despite definitional debates, 'open-weights' AI models released publicly by Meta (Llama series), Mistral AI, and others created a vibrant ecosystem of researchers, developers, and companies building on top of pre-trained foundations without needing to train from scratch. DeepSeek's R1 model, released in January 2025, demonstrated that open-weight models could achieve performance competitive with the leading proprietary frontier models at a fraction of the training cost, reshaping competitive dynamics across the AI industry. Example: DeepSeek released its R1 reasoning model in January 2025 with full weights and a permissive license, demonstrating near-frontier performance on mathematics and coding benchmarks while reportedly being trained for significantly less compute cost than comparable proprietary models. R1's release triggered widespread discussion about the sustainability of the moats around proprietary AI labs and accelerated open-source AI adoption across enterprise applications that had previously avoided frontier AI due to data privacy concerns. Why it matters for AI: Open-source AI is critical to the democratization of AI development, enabling researchers without access to hyperscaler compute, companies with data privacy requirements that prohibit sending data to third-party APIs, and developers in lower-income countries to access frontier AI capabilities. It also serves as a check on the concentration of AI capabilities in a small number of proprietary labs, creating competitive pressure and enabling independent safety research on the same model architectures used in production.

Explore the full Web3 Glossary — 2,000+ expert-curated definitions. Need guidance? Talk to our consultants.