EXL3 quants of Qwen3.5-122B-A10B

⚠️ Requires ExLlamaV3 v0.0.23 (or v0.0.22 dev branch)

Base bitrates:

2.00 bits per weight
3.00 bits per weight
4.00 bits per weight
5.00 bits per weight

Optimized:

2.28 bits per weight
3.06 bits per weight
4.06 bits per weight

Benchmarks

. Ppl¹ KL-div HumanEval @1²
2.00 bpw 7.672 0.594 87.80%
2.28 bpw 5.077 0.243 92.07%
3.00 bpw 4.818 0.146 94.51%
3.06 bpw 4.487 0.100 95.12%
4.00 bpw 4.312 0.058 94.51%
4.06 bpw 4.219 0.042 95.12%
5.00 bpw 4.177 0.031 95.12%
Original 4.264

¹ (10 rows of wikitext2) ² Reasoning disabled

(more coming)