What could change the answer
- Long context and KV cache
- Batch size or concurrent users
- Runtime overhead and driver support
- CPU/RAM offload
- Image encoders and auxiliary models
MODEL × HARDWARECan L40S 48GB
Can L40S 48GB
run Gemma 3 12B?
Yes, with headroom.
Q4_K_M MEMORY VERDICTCOMFORTABLE
9.9 GB estimated requirement vs 48.0 GB usable-memory assumption.
Model weights7.3 GB
Runtime reserveIncluded
Hardware stackCUDA
Speed claimNot estimated