Weights

How it works

A model is an architecture (the code that defines its layers) plus weights (the trained numbers). Weights are distributed as one or more files in formats such as safetensors or GGUF, alongside a configuration file and the tokenizer files.

For closed models the weights stay with the provider, and you reach the model only through an API. For open-weights models you download the files under a license that may restrict use; open weights does not mean the training data or training code is open. One release can include several checkpoints, such as base and instruct versions, different sizes or different precisions.

Weights are fixed after training, so a model does not change as you use it. Behaviour changes come from a new checkpoint, fine-tuning or adapters such as LoRA. Download from the maker's official repository and verify the files, because re-uploads can be modified or mislabeled.

Worked example

Illustrative: a 12B-parameter model at BF16 is about 12 x 2 = 24 GB of weight files, often split into shards of about 5 GB, so five files. A 4-bit version is about 12 x 0.5 = 6 GB plus some overhead for scales.

Common mistakes

• 1Treating open weights as open source; training data, code and license terms are separate questions.
• 2Downloading weights from unofficial re-uploads without checking they match the maker's release.
• 3Loading pickle-based weight files from untrusted sources, which can run code; prefer safetensors.