The AI industry is using "open source" terminology to describe models that provide only a fraction of what the term has historically meant, according to Peter Farkas, co-creator of open source MongoDB alternative FerretDB and CEO of database services company Percona. Speaking at Open Source Summit Europe in Prague, Farkas called on developers not to treat "open weight" and "open source" as equivalent terms, arguing that conflating the two threatens to erode the fundamental freedoms associated with open source software. The warning comes as open-weight models have become a major force in production AI environments.
In August, open-weight models accounted for 56% of tokens processed through Vercel's AI Gateway and 60% of US-originating token consumption on OpenRouter, with Chinese-developed models making up the majority. An AI model's weights are the numerical parameters created during training that capture the patterns the model has learned. Releasing these weights allows developers to download and operate a model on their own infrastructure, even when the creator hasn't shared the ingredients or process used to build it. James Landay, director of Stanford's Institute for Human-Centered AI, drew a distinction in an August article: open weights answer "can I run this?" while open source answers "can I trust this, improve it, and build the next thing on top of it?"
Farkas pointed to China's DeepSeek as an example of the confusion, noting its models are widely called open source despite the company releasing model weights rather than everything needed to reproduce the models from scratch. However, he emphasized that downloading and running those models remains valuable on its own. "Are open weights a bad thing? No, open weights are great," Farkas said. "The problem is when open weights are positioned as, 'hey, this is as good as open source'." Some companies are going further: Xiaomi recently launched its MiMo-V2.6 models and livestreamed nearly a week of reinforcement-learning training through a public dashboard, while also releasing more than 7,000 reinforcement-learning task environments along with its RL training code and technical documentation.
Without a commonly understood boundary, Farkas argued, "open washing wins". He expressed concern that accepting a looser definition of open source for AI could ultimately weaken the definition for software more broadly, warning that companies might call something Apache 2.0 that can't be used in the European Union. The Open Source Initiative published its first Open Source AI Definition in 2024, establishing criteria around the freedoms to use, study, modify, and share AI systems, but the definition remains contested particularly over what information about training data an AI system must disclose. Responding from the audience, OSI's new executive director Duane O'Brien acknowledged the criticisms and confirmed the organization is reopening the discussion, recently launching an Open Source AI Fellowship with Gabriel Toscano as its first fellow to help build consensus. The OSI plans a series of community discussions over the next two years to explore potential revisions to the definition. The terminology battle reflects a deeper tension between technical accessibility and the collaborative transparency that has defined open source culture for decades, with the stakes extending far beyond any single industry's vocabulary.