TurboQuant · TQ3_4S · single RTX 3090

Laguna XS 2.1-TQ3_4S
3B active. Fastest here.

100%coding suite

96.4%speed suite

3Bactive · 30B MoE

4.05bits / weight

196tok/s

decode · greedy

1389tok/s

pp128 · prefill

Model Decodetok/s · greedy HumanEval+pass@1 MBPP+pass@1 Hard8686 assertions

Laguna XS 2.1

TQ3_4S · 4.05 bpw · RTX 3090

196

1389pp128

76.2%

72.0%

64.0%

Ornith 35B

TQ3_4S · 13.0 GiB · RTX 3090

146.3

87.2%

70.9%

81.4%

Qwen3.5 9B

TQ3_4S · RTX 3090

134

67.1%

56.3%

44.2%

Qwen3.6 27B MTP

TQ3_4S · RTX 3090

92.7%*

87.8%*

88.4%