Not long ago, the best language models were available only through a handful of proprietary APIs. Today, open-weight models that anyone can download and run are within striking distance on many everyday tasks, and the implications reach far beyond benchmarks.
What "open" really means
Most so-called open models are open-weight: you can download and run them, but the training data and full recipe are not published. Licences vary widely, from permissive to restrictive for large commercial users. Read the licence before you build a product on one.
Why companies are paying attention
Running a model on your own infrastructure means sensitive data never leaves your network, costs become predictable, and you can fine-tune on proprietary data. For high-volume, narrow tasks, a smaller open model often beats a frontier API on cost by an order of magnitude.
Where closed models still lead
The very hardest reasoning, coding and multimodal tasks still favour frontier systems, and managed APIs remove a large operational burden. Most mature teams now run a mix: a frontier model for difficult requests and an efficient open model for the bulk of traffic.
Key takeaways
- What "open" really means
- Why companies are paying attention
- Where closed models still lead
