The estimate above is a per-token bill from a provider's API. A model with downloadable weights can also run on hardware you own or rent, where you pay for the machine by the hour and not for tokens. That can cost less at steady, high volume and more at low or bursty volume, and it adds operations work. Quality also differs between models, so compare on your own tasks.
This page does not estimate a hosting bill. To compare, use Compare this workload with GPU hosting in the result above, which carries your request sizes and volume to a hosting calculator, and check that the model fits your hardware with the VRAM calculator.
Guides: Self-hosting LLMs covers serving a model for a team or a product, and How to run LLMs locally covers running one on your own computer.
Newest release first, not a quality ranking. "Custom license" means commercial use is allowed with conditions: read the license before shipping. Sizes are the weights alone at 4-bit quantization; context and runtime overhead add to that. All open-weight language models