Qwen3.8-27B-Uncensored — AWQ 5.0bpw (gs64)

The quantization recipe of True2456/Qwen3.8-27B-AWQ-5.0bpw applied to orcarouter/Qwen3.8-27B-Uncensored.

17.35 GB · BPW 5.00 · same per-module bit map as the donor (MLP 4-bit gs64, GDN in_proj 5-bit gs64, attention q/k/v 8-bit gs64, o_proj 4-bit gs64, lm_head 6-bit gs128, embeddings 4-bit gs128, vision tower 8-bit, MTP head 8/6/4-bit).

How it was made

Per-module {bits, group_size} map taken verbatim from the donor's config.json (mode: affine).
The donor's AWQ calibration was transplanted, not re-run: the donor folds a per-hidden-channel scale s into each post_attention_layernorm and a per-intermediate-channel scale t into up_proj rows / down_proj columns. Both were recovered from the donor (the fine-tune's norms are bit-identical to base Qwen, which makes s exact; t was recovered from the donor's dequantized down_proj by robust median ratio). The same folds were applied to the fine-tune's weights before quantization — mathematically an identity transform, so nothing of the fine-tune is lost.
On top of the transplanted scales, LM MLP weights got a per-group MSE clip search (α grid 1.0→0.90), which measures slightly better weight-MSE than the donor build (0.0892 vs 0.0913 rel RMS on layer 0 gate_proj).
Everything outside the LM MLP is RTN at the donor's widths — verified bit-exact against the donor on the MTP head and all post_attention_layernorm tensors (64/64). The only intentional deviations are mtp.layers.0.self_attn.o_proj and mtp.layers.0.mlp.down_proj, which in the donor carry a double-quantization artifact from its 2026-08-16 MTP repair; here they are quantized directly from bf16 (slightly lower error).
MTP head ships inside the checkpoint (31 tensors under mtp.*), quantized 8/6/4-bit like the donor. The 27 vision linear_fc2 stay bf16 (input dim 4304 not divisible by any supported group size), same as the donor.

Requires oMLX

Same caveat as the donor: the checkpoint keeps mtp.* weights so the MTP head can bind. Stock mlx_lm/mlx_vlm either drop MTP + vision or mis-apply the norm shift. Load with oMLX (mtp_enabled for native multi-token-prediction decode).

Tokenizer, chat template and configs are the fine-tune's own files.

Qwen3.8-27B-Uncensored — AWQ 5.0bpw (gs64)

How it was made

Per-module {bits, group_size} map taken verbatim from the donor's config.json (mode: affine).

The donor's AWQ calibration was transplanted, not re-run: the donor folds a per-hidden-channel scale s into each post_attention_layernorm and a per-intermediate-channel scale t into up_proj rows / down_proj columns. Both were recovered from the donor (the fine-tune's norms are bit-identical to base Qwen, which makes s exact; t was recovered from the donor's dequantized down_proj by robust median ratio). The same folds were applied to the fine-tune's weights before quantization — mathematically an identity transform, so nothing of the fine-tune is lost.

On top of the transplanted scales, LM MLP weights got a per-group MSE clip search (α grid 1.0→0.90), which measures slightly better weight-MSE than the donor build (0.0892 vs 0.0913 rel RMS on layer 0 gate_proj).

Everything outside the LM MLP is RTN at the donor's widths — verified bit-exact against the donor on the MTP head and all post_attention_layernorm tensors (64/64). The only intentional deviations are mtp.layers.0.self_attn.o_proj and mtp.layers.0.mlp.down_proj, which in the donor carry a double-quantization artifact from its 2026-08-16 MTP repair; here they are quantized directly from bf16 (slightly lower error).

MTP head ships inside the checkpoint (31 tensors under mtp.*), quantized 8/6/4-bit like the donor. The 27 vision linear_fc2 stay bf16 (input dim 4304 not divisible by any supported group size), same as the donor.