Foundations & Models
Open-Weight Model
The weights are available for download and the model runs on your own hardware: full data sovereignty, no API dependency, adaptable through fine-tuning. But open does not automatically mean transparent — the training data usually stays secret, and the license may restrict use. For sensitive data and predictable costs, open weight is often the most sober choice.
An example
A law firm wants briefs summarized. Client data must not leave the firm, so a cloud service is out of the question. With an open-weight model, the analysis runs on a machine in the firm's own server room: the data stays where it is, there is no data processor, no transfer to a third country and no question about use for training. The price is operating effort — someone has to maintain the machine, keep the model current and measure quality.
A closer look
Freely available means downloadable and executable locally, usually under a license with conditions: usage limits above a certain company size, bans on particular uses. It is worth reading the license; it differs considerably from provider to provider.
Why it matters
“Open weight” is not the same as open source. Usually only the result is published, not the training dataset and not the training code; the model cannot be rebuilt or fully audited with that. The licenses also vary in strictness — some prohibit commercial use or certain applications. Anyone who relies on open weights for compliance reasons should read the license as carefully as a cloud contract. The freedom is real, but it is not unlimited.
Commonly misunderstood
Running it yourself is not automatically cheaper. Below a certain utilization, a reserved graphics card costs more than the same number of requests through an API.