Open-weight artificial intelligence models, whose underlying settings are released for anyone to download and run rather than locked behind a company's paywall, have passed a milestone in developer usage.

They made up 53% of token volume on Vercel's AI Gateway on 25 August, according to a new note from Citi, the US investment bank.

The gateway is a service that lets software developers plug into many different AI models, while tokens are the small chunks of text those models read and generate, so token volume roughly tracks how much each type of model is being used.

That 53% share was up 24 percentage points from late June, Citi's Inference Ahead report said.

Over the same period, the performance gap between the strongest proprietary models and the strongest open-weight ones narrowed from 9 points to 3.

Citi said publicly available frontier models, meaning the most advanced systems on the market, are being held back by regulatory constraints.

The commercial case for open models is also improving. DeepSeek, the Chinese developer behind a low-cost system, has reportedly generated $71 million in revenue through July, around ten times its full-year 2025 total, at an 83% gross margin on paid access to its model.

The American investment bank said such numbers show open-weight models can support attractive economics.

Nvidia, the chipmaker at the heart of the AI boom, has reportedly agreed to buy Hugging Face, a widely used hub for open models, for $12.9 billion, and to license technology from the developer Poolside for $6 billion.

Citi said those deals could strengthen the open-weight ecosystem while driving demand for computing power.

The bank pointed to next week's LEAP technology event as an early read on how wider open-model adoption is affecting corporate AI deployment and infrastructure spending.