120b hits near parity with o4-mini on core reasoning and runs on a single 80 GB GPU.

20b lands around o3-mini on common benchmarks and runs in 16 GB.