Mia-AiLab/Qwen3.8-27B-EXL3

Mirror notice: This repository is a community mirror of turboderp/Qwen3.8-27B-exl3. These are not my weights. They are copied from turboderp's originals so the files stay available if upstream copies are removed for any reason. All credit and ownership remain with turboderp. Mirrored on 2026-09-03.

EXL3 quants of Qwen3.8-27B

Calibration trace: md JSON safetensors
Eval trace: md JSON

Self-calibrated quants:

Rightmost column are variants with quantized vision tower and require ExLlamaV3 v1.4.4 to work properly

1.40 bits per weight / H3 -- 1.40 bits per weight / H3 / V3
1.60 bits per weight / H3 -- 1.60 bits per weight / H3 / V3
1.80 bits per weight / H3 -- 1.80 bits per weight / H3 / V3
2.00 bits per weight / H3 -- 2.00 bits per weight / H3 / V3
2.20 bits per weight / H3 -- 2.20 bits per weight / H3 / V3
3.00 bits per weight / H4 -- 3.00 bits per weight / H4 / V4
4.00 bits per weight / H5 -- 4.00 bits per weight / H5 / V6
5.00 bits per weight / H6 -- 5.00 bits per weight / H6 / V6
6.00 bits per weight / H6 -- 6.00 bits per weight / H6 / V6

Plain quants:

2.00 bits per weight
2.50 bits per weight
3.00 bits per weight
3.50 bits per weight
4.00 bits per weight
5.00 bits per weight
6.00 bits per weight