community-verified · measured pool data

Can the RTX 3090 24GB run Phi-4-mini-reasoning?

Yes — community-measured at 110.6 tok/s (9 runs at 4-bit).

Verified single-stream runs on the RTX 3090 24GB from the localmaxxing public pool. Every number on this page is measured — nothing here is estimated.

Quantization Median tok/s TTFT (ms) Peak VRAM Max context tested Engines
4-bit 110.6 1403.4 0 GB 131,072 llama.cpp
Go deeper
I have a goal — find models that fit my use case → I have hardware — browse everything my rig can run → All machine × model answer pages →