Open weights

How it works

An open-weights model has its trained parameters published for download, so you can run it on your own hardware, fine-tune it and inspect it. This is not the same as fully open source: training data and training code are often not released.

Releases differ in license terms, available sizes, whether base and instruction-tuned versions are both published and which precisions are offered. Many are also served per token by third-party hosts, so you can use them without running GPUs.

Open weights give you control over where data goes, which version runs and how the model is customized, and nobody can retire your copy. You take on serving, scaling, security updates and the safety filtering an API provider would otherwise handle.

Worked example

Illustrative: a 70B-parameter model at 16-bit precision needs about 70B x 2 bytes = 140 GB of memory for the weights alone, more than one 80 GB GPU holds. At 4-bit it is about 35 GB.

Common mistakes

• 1Treating 'open weights' as 'open source' and assuming any use is allowed without reading the license.
• 2Forgetting that safety filtering and abuse monitoring are now your job.
• 3Sizing hardware for the weights only, not for the memory each concurrent request needs.