Open Source Models and the Open Ecosystem

An "open" AI model is usually not open source in the software sense. What most labs release is open weights: the trained parameters, which anyone can download, run, fine-tune and often sell services on, under a licence that may or may not resemble a software open-source licence. The training data and much of the training code usually stay private. Open weights changed the economics of the whole stack. Each time a strong open model appears (Meta's Llama 2 in July 2023, Mistral's Mixtral in December 2023, Llama 3.1 405B in July 2024, DeepSeek-R1 in January 2025, Alibaba's Qwen3 in April 2025, OpenAI's gpt-oss in August 2025), the price that closed labs can charge for comparable capability falls, application builders gain a fallback against platform risk, and clouds, chip makers and inference hosts gain a product they can sell without paying a lab. But open weights do not pay for themselves: the releasing lab spends tens or hundreds of millions of dollars and gives away the result. An open-model strategy is therefore an answer to four questions: what does the releaser sell instead of the model, how open is the release really, who captures the value it creates, and how fast does the open frontier follow the closed one? This section explains the spectrum of openness, the licences, why each kind of releaser gives models away, the hub, hosting and fine-tuning economy around open weights (Hugging Face, inference clouds, local runtimes), the commoditisation pressure on closed labs, the shift of open leadership from the US to China in 2025–26, and the policy fights over whether open weights are a security risk or a public good.