On an RTX 5070 Ti (16 GB), Gemma 3 12B in Q4_K_M needs about 8,3 Go of video memory: it should fit. Estimated generation speed: ≈ 106 tokens/s.
Not measured yet: the figures below are estimates, computed from the models already measured on this card. Measure my card
≈ 106 tok/s
Generation (estimated)
≥ 8,3 Go
Recommended VRAM (16 GB avail.)
≈ 152 W
Power (ref. model)
0,10 €/M tok
Electricity / M tokens · 0,25 €/kWh
Calculation: Qwen3.5 9B measured at 125 tok/s on this card, 5,7 Go read per token → × 5,7 Go ÷ 6,8 Go ≈ 106 tok/s (generation is memory-bound: speed follows the amount of weights read).
What about a longer context?
Context
KV cache
Estimated VRAM
Fits in 16 GB?
Estimated speed
4 096
0,3 Go
7,5 Go
yes
≈ 103 tok/s
8 192
0,5 Go
7,8 Go
yes
≈ 99,3 tok/s
16 384
1,0 Go
8,3 Go
yes
≈ 92,9 tok/s
32 768
2,0 Go
9,3 Go
yes
≈ 82,3 tok/s
65 536
4,0 Go
11,3 Go
yes
≈ 67,0 tok/s
131 072
8,0 Go
15,3 Go
yes
≈ 48,8 tok/s
Estimated VRAM = file + KV cache (full-attention layers, f16) + ~0,5 Go of buffers, excluding memory used by the system — hence the slightly higher “recommended VRAM”, which keeps that margin. Speed ≈ estimated speed × weights read ÷ (weights read + added KV cache).
Estimates based on real measurements of other models on this card and on the model's public data (Google Gemma). As soon as a measurement exists, this page shows the real figures.