Open weights

Open weights are a model's trained parameters, published so anyone can download, run, and modify them. That contrasts with closed models, where the provider keeps the weights and sells access through an API instead. Meta's Llama and Alibaba's Qwen both ship this way, usually as safetensors or GGUF files hosted on Hugging Face.

Parameter count sets the baseline file size. A 70 billion parameter model at 16-bit precision needs about 140 GB of weight files. Quantized to 4-bit that drops to roughly 35 to 40 GB, which fits a single 48 GB workstation card. Licenses vary even when the files are freely downloadable. Some releases use permissive terms like Apache 2.0, and others cap free commercial use above a set user count.

Size memory to the weight file you actually plan to run, not the headline parameter count. The same 70B label can mean 140 GB at 16-bit or under 40 GB quantized. That gap decides whether it fits on one GPU or needs several.

Sources

Last verified August 29, 2026.