Every local LLM on CanRun

All models

The memory column is for loading the whole model on one graphics card. To see what runs on your own PC, Mac or laptop, use the calculator. Check my PC in the calculator · Browse GPUs, Macs and PCs

Popular first follows a five-level score CanRun gives each model, so it can differ from the downloads column. Pick “Most downloaded” to sort by downloads alone.

Scroll sideways to see every column.

How to read this list

• Hugging Face org: the organization that publishes the original weights, under the name it uses on Hugging Face. Some of these are brand names rather than company names.
• Parameters: the total count. Mixture-of-experts (MoE) models also show their active count. The whole model has to fit in memory, but each token reads only the active part, so an MoE model usually generates faster than a dense model of the same total size. The size filter groups MoE models by their total.
• Memory needed: graphics card memory for the Q4_K_M file with 8k context. Models without a Q4_K_M file use the closest quant, named in brackets. It adds up the weights, the KV cache and the compute buffer, the same calculation as the memory table on each model page. Add about 0.6 GB if the GPU also drives your display on Windows.
• Released: the date the original repository was created on Hugging Face. A repository prepared privately can predate the public release.
• Thinking mode: the model writes out its reasoning before it answers. That costs extra tokens, so its verdicts use speed thresholds 1.5× stricter. “Always on” models always think, “On by default” and “Off by default” models can be switched, and “Model decides” models choose for themselves.
• License: Open means a permissive license such as Apache-2.0 or MIT. Custom terms means the maker’s own license, so read it before commercial use. Non-commercial licenses do not allow commercial use.
• Downloads: Hugging Face downloads of the original repository over the last 30 days, as of October 21, 2026. Downloads of GGUF copies are not counted, so the number can understate how widely a model is used.
• Popular first follows a five-level score CanRun gives each model, weighing downloads together with how widely it is used in Korea. Models with the same score are ordered by downloads.