Groot-19B is a free Apache 2.0 model its makers say runs on a 16 GB GPU; we ran it with no GPU at all
Groot-19B-A3B-261023 from GrootTeam is a free, open-weight (Apache 2.0) mixture-of-experts model: about 19B total parameters, with about 3B active per token. It is post-trained from Qwen3.6-35B-A3B. Its makers say it beats Qwen3.6-35B-A3B and Gemma-4-26B-A4B on 12 of 16 benchmarks. Those are the makers' own benchmarks, so we tested it. The official Q4_K_M GGUF is an 11.7 GB download, and its makers say it runs on a 16 GB card like an RTX 4060 Ti. On our box with no GPU, just 8 CPU cores, it answered our question correctly at 7.9 tokens/s.
Try it
llama-server -hf GrootTeam/Groot-19B-A3B-261023-GGUF:Q4_K_MNeeds llama.cpp; 11.7 GB download. Then open localhost:8080. We ran the same -hf download with llama-cli (llama.cpp b11960) on a CPU-only Linux box with thinking turned off; we did not test on a GPU.