On an RTX 5070 Ti (16 GB), Ministral 3 3B in Q4_K_M needs about 3,5 Go of video memory: it should fit. Estimated generation speed: ≈ 266 tokens/s.
Not measured yet: the figures below are estimates, computed from the models already measured on this card. Measure my card
≈ 266 tok/s
Generation (estimated)
≥ 3,5 Go
Recommended VRAM (16 GB avail.)
≈ 121 W
Power (ref. model)
0,03 €/M tok
Electricity / M tokens · 0,25 €/kWh
Calculation: Llama 3.2 3B Instruct measured at 283 tok/s on this card, 1,9 Go read per token → × 1,9 Go ÷ 2,0 Go ≈ 266 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
0,4 Go
2,9 Go
yes
≈ 227 tok/s
8 192
0,8 Go
3,3 Go
yes
≈ 194 tok/s
16 384
1,6 Go
4,1 Go
yes
≈ 150 tok/s
32 768
3,3 Go
5,7 Go
yes
≈ 103 tok/s
65 536
6,5 Go
9,0 Go
yes
≈ 63,1 tok/s
131 072
13,0 Go
15,5 Go
yes
≈ 35,7 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 (Mistral). As soon as a measurement exists, this page shows the real figures.