Open weights
open-weight model · downloadable model
กล่าวโดยย่อ
Open weights means the trained parameter file is published for download, so you can run, inspect and fine-tune the model on your own hardware. The licence attached decides what you may legally do with it, and many open-weight models ship under custom licences that do not meet the open source definition.
Downloadable weights buy four things a hosted API cannot: data residency, because text never leaves your infrastructure; version stability, because the model does not change underneath your evaluations; cost structure, trading per-token pricing for hardware you control; and inspectability, since you can probe, quantise, prune and fine-tune freely.
They also transfer work to you — serving infrastructure, utilisation, monitoring, upgrades, and the safety filtering a provider would otherwise supply.
The licensing distinction is where teams get caught. "Open" in an announcement can mean an OSI-approved licence such as Apache-2.0, or a bespoke community licence carrying an acceptable-use policy, a monthly-active-user threshold above which a separate agreement is required, a naming requirement for derivatives, or restrictions on using outputs to train competing models.
None of those terms is unreasonable. They are simply not open source, and they matter concretely for white-labelling, procurement review and reselling a fine-tuned derivative. Read the licence file that ships with the weights rather than the blog post announcing them — see open weights vs open source.
คำถามที่พบบ่อย
- Is open weights the same as open source?
- No. Open source is a licence standard requiring freedom to use, study, modify and redistribute without field-of-use restrictions. Several popular open-weight models carry custom licences with acceptable-use policies or user thresholds attached.
- Can I see the training data of an open-weight model?
- Almost never. You get the parameters and usually a model card describing the data in general terms, not the corpus itself.