On an RTX 5070 Ti (16 GB), Gemma 4 12B in Q4_K_M needs about 8,6 Go of video memory: it should fit. Estimated generation speed: ≈ 95,6 tokens/s.
Not measured yet: the figures below are estimates, computed from the models already measured on this card. Measure my card
≈ 95,6 tok/s
Generation (estimated)
≥ 8,6 Go
Recommended VRAM (16 GB avail.)
≈ 179 W
Power (ref. model)
0,13 €/M tok
Electricity / M tokens · 0,25 €/kWh
Calculation: DeepSeek R1 Distill Qwen 14B measured at 81,5 tok/s on this card, 8,4 Go read per token → × 8,4 Go ÷ 7,1 Go ≈ 95,6 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
7,9 Go
yes
≈ 92,9 tok/s
8 192
0,5 Go
8,1 Go
yes
≈ 89,8 tok/s
16 384
1,0 Go
8,6 Go
yes
≈ 84,3 tok/s
32 768
2,0 Go
9,6 Go
yes
≈ 75,0 tok/s
65 536
4,0 Go
11,6 Go
yes
≈ 61,5 tok/s
131 072
8,0 Go
15,6 Go
yes
≈ 45,2 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.