whichllm is a useful Python CLI utility that autodetects your GPU/CPU/RAM and pulls benchmarks from Hugging Face to rank the local LLMs you can run on your system, based on a mix of benchmarks and projected token speeds on your machine.
I have a modest but respectable 16 GB of NVIDIA VRAM at my disposal, so unsurprisingly the top models it suggests for my hardware are in the Qwen 3 and Phi 4 families, as well as the always excellent gpt-oss-20b from OpenAI. These suggestions basically match my own experiments.
The top suggestion is Qwen3.6-27B, but the tool is basing it on stale benchmarks from May 2026 (I got a rate limiting error when it tried to pull the live benchmarks—I guess a lot of people are clamoring to test local LLMs).
If the tool was using live benchmarks, I believe it would almost certainly put Qwen3.8-27B at the very top, the model released a few weeks ago to rave reviews.
I was also curious which models it would recommend if I was running on CPU only. Surprisingly, the top suggestion was gpt-oss-20b, with a warning that all of the top contenders would run very slowly and that token speed estimates were uncertain.