Open-weight models and the economics of open source
Open-source software appears on 13 recent papers and open-source language models on 6, while named open-weight models such as DeepSeek, Llama and Qwen appear on dozens of AI papers as tools; 78 of 126 open-source papers involve language models. Firms give away models worth hundreds of millions to train. Who adopts them, and what do the givers get?
What prompted this
Since October 2025, 126 papers carry an open-source, open-weight or named open model keyword (96 in AI journals, 13 in economics, 8 in management): "DeepSeek" on 12, "DeepSeek-R1" on 11, "Llama 3" on 10, "Qwen2.5" on 6, "open source software" on 13, "open-source LLMs" on 6. Seventy-eight of the 126 also carry a language-model, foundation-model or fine-tuning keyword. The economics of open source, on 41 papers with 9 touching firms or investment, has not caught up with models.
Questions
- Which firms deploy open-weight models rather than paid APIs, and do they do so for cost, privacy, control or regulation?
- Do the firms that release open models gain in cloud revenue, hiring, standards or valuation, and did the release of strong open models from China change the strategy of US firms?
- Does the availability of open models lower prices for closed ones, as open-source software did for proprietary software?
Data that could help
Model download and fine-tuning statistics (Hugging Face), job postings naming models, API price histories and firm disclosures.
Next step
Track API prices of closed models around the release dates of major open-weight models and measure the price response.