Open models quietly became the enterprise default
Somewhere in the last year, open-weight models stopped being a science project. By mid-2026 they are an enterprise standard, with open models like Llama holding a large share of business deployments and a fast-growing catalog of strong alternatives behind them.
The more interesting shift is downward. Small models keep getting better: a sub-4-billion-parameter model can now reach a large fraction of frontier performance and run directly on a laptop, with no data leaving the device. For the bulk of business work, that is already more than enough.
The Stavryn take
- You can now hold genuinely capable AI inside your own building, on hardware you own, without renting a frontier model by the token.
- Right-sized and private beats biggest-and-rented for most real work.
See the case against frontier models, how we build it, or start with what private AI is.