Decode · tok/s
54.7
Prefill
—
tok/100W
30.4
I tested Ternary Bonsai 2 27B from PrismML on my RTX 5060 Ti 16 GB. It decodes 2x faster than a normal 4-bit quant and needs half the VRAM. Bonsai PTQ1_0 vs Qwen3.8-27B IQ4_XS, same card, same settings, thinking off: Decode: 26.6 → 54.7 tok/s VRAM: 13.6 → 6.3 GB
First-person RTX 5060 Ti 16GB. Rank is the quoted 54.7 decode for Bonsai PTQ1_0 thinking off. 26.6 is the IQ4_XS companion on the same card, not a separate rank. Runtime not named in the post; set to llama.cpp from GGUF quant style.
HF GGUF