On an RTX 5070 Ti (16 GB), R1 Distill 8B Llama in Q4_K_M needs about 6,1 Go of video memory: it should fit. Estimated generation speed: ≈ 151 tokens/s.
Not measured yet: the figures below are estimates, computed from the models already measured on this card. Measure my card
≈ 151 tok/s
Generation (estimated)
≥ 6,1 Go
Recommended VRAM (16 GB avail.)
≈ 148 W
Power (ref. model)
0,07 €/M tok
Electricity / M tokens · 0,25 €/kWh
Calculation: Meta Llama 3.1 8B Instruct measured at 151 tok/s on this card, 4,6 Go read per token → × 4,6 Go ÷ 4,6 Go ≈ 151 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
0,5 Go
5,6 Go
yes
≈ 138 tok/s
8 192
1,0 Go
6,1 Go
yes
≈ 125 tok/s
16 384
2,0 Go
7,1 Go
yes
≈ 106 tok/s
32 768
4,0 Go
9,1 Go
yes
≈ 81,1 tok/s
65 536
8,0 Go
13,1 Go
yes
≈ 55,2 tok/s
131 072
16,0 Go
21,1 Go
no
—
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 (DeepSeek). As soon as a measurement exists, this page shows the real figures.