At the Open Source Summit Europe in Prague this week, Percona chief executive Peter Farkas made a request that sounds pedantic and turns out not to be: "Don't use 'open weight' and 'open source' interchangeably." His point is that the two words promise very different things. Releasing a model's weights, the numerical parameters produced by training, lets you download and run it. It does not hand over the source code or the training data, the ingredients and recipe that would let you understand, reproduce or truly build on it.
Stanford's James Landay has drawn the line more memorably. "Open weights answer 'can I run this?'" he has said. "Open source answers 'can I trust this, improve it, and build the next thing on top of it?'" By that test, most of today's celebrated "open" models answer only the first question. DeepSeek is routinely called open source despite releasing weights rather than everything needed to rebuild it from scratch. Farkas's worry is "open washing": if the industry lets the looser meaning stick, companies could slap an Apache 2.0 label on something you cannot actually use freely, and the erosion would spread from AI back into software generally.
This would be a dry definitional spat if the terms were not quietly rewiring the market. Open-weight models are no longer a hobbyist corner. In August they accounted for 56 percent of tokens flowing through Vercel's AI Gateway and 60 percent of US-originating consumption on OpenRouter, and Chinese-developed models made up the majority of that. Businesses are reaching for them because they are cheap and controllable: run the model on your own hardware, keep sensitive data in house, and avoid lock-in. A JPMorgan analysis cited this week found some Chinese models cost 10 to 50 times less per token than leading proprietary systems.
That price gap is reshaping the US-China contest in a way neither government planned. As analyst Christopher McNally argues, Chinese labs, boxed out of the best chips by export controls, leaned hard into open weights as a fast-follower strategy, and it worked: their models have surged past American ones on Hugging Face, capturing roughly 41 percent of platform supply. The same openness that spreads capability also spreads risk. Strip an open-weight model's guardrails, and a safe frontier system can be fine-tuned into a malicious one. The feature and the vulnerability are the same feature.
Into this gap steps Europe, trying to make a virtue of sovereignty. Mistral is pitching its latest open-weight model as a way for governments and companies to keep advanced AI on their own infrastructure rather than handing workloads to a US or Chinese provider, a case strengthened by a fresh funding round of more than 3.3 billion dollars. Whether "built in Europe" becomes a meaningful third option or just another label is the open question.
The standards bodies are scrambling to catch up. The Open Source Initiative, whose first Open Source AI Definition in 2024 drew fire for being too permissive about training-data disclosure, has conceded the criticisms were valid and reopened the debate, appointing a fellow to rebuild consensus over two years. Two years is a long time in this field. By the time the definition settles, the market may already have decided what "open" means in practice, and the people writing the dictionary will be describing a world rather than shaping it.