LOCAL AIregistry
‹ all GPUs

RTX 5070

NVIDIA · 12 GB · 672 GB/s
The recipe
Recommended
Qwen3.5-9B
EXL3 4 bpwTabbyAPI128K context

Best for Fast answers and light agent work on 8–12 GB cards.

129 tok/sdecode
1,964 tok/sprefill
128Kcontext
✓ Thinks✓ Tools✓ Long context
Tested on this card · Sep 25, 2026