Apple MLX quant of Qwen/Qwen3.8-27B (text tower, qwen3_5 hybrid arch: transformer + Gated-DeltaNet + MTP) using the per-layer bit allocation of Unsloth's Dynamic v3.0 Q3_K_XL GGUF, applied to clean fp16 weights (no double quantization).

3.95 bits/weight · 12.4 GB · loads and runs in oMLX / mlx-lm 0.31.3+

Why this model

The quant recipe is not a fixed bit-width: each tensor keeps the bit-width the unsloth v3.0 imatrix calibration chose for it in the Q3_K_XL GGUF (2/3-bit where robust, 5/6/8-bit where sensitive), re-encoded in MLX's affine group-quant format.
It was quantized once, from the fp16 checkpoint — not derived from the already-quantized GGUF, so there is no compounding quantization error.

Bit allocation (per tensor, group_size 64, affine)

Bits	Tensors	What
2	24	robust GDN/attention projections (from Q2/IQ2 source layers)
3	157	MLP gate/up, GDN in-proj (from Q3/IQ3 source layers)
4	194	bulk MLP down, attention, GDN (from Q4/IQ4 source layers)
5	26	sensitive projections, lm_head (from Q5_K source layers)
6	1	lm_head (from Q6_K source layer)
8	96	conv1d + high-value projections (from Q8_0 source layers)
bf16	—	norms, A_log, dt_bias, F16/F32 source tensors

Verification (M1 Pro 32 GB, wikitext-2, 64×512-token samples, seed 123)

Perplexity: 12.724 ± 0.212
Decode: 10.17 t/s, prefill 49.2 t/s (batch 1)
Peak memory: 13.8 GB

For reference, the same checkpoint with the published Unsloth Dynamic 2.0 static per-tensor table (5.27 bpw, 17 GB) scores 13.445 ± 0.231 at 8.23 t/s. The v3.0 allocation is strictly better on every axis at 4.6 GB smaller.

Usage

bash

pip install mlx-lm
python -m mlx_lm.generate --model jclyons52/Qwen3.8-27B-Unsloth-Dynamic-v3-mlx \
  --prompt "Explain quantization in one sentence."

Serve:

python -m mlx_lm.server --model jclyons52/Qwen3.8-27B-Unsloth-Dynamic-v3-mlx

oMLX: add the repo id jclyons52/Qwen3.8-27B-Unsloth-Dynamic-v3-mlx.

How it was built

Downloaded Qwen/Qwen3.8-27B fp16 (18 shards, ~54 GB).
Parsed the official unsloth/Qwen3.8-27B-GGUF Q3_K_XL file's per-tensor quant types (IQ2/IQ3/IQ4/IQ5/Q6/Q8 mixes) and mapped each to an MLX affine bit-width.
mlx_lm.convert with that per-tensor predicate, dtype=bfloat16, group_size=64 — text-only (vision tower and MTP head are not included; mlx-lm 0.31.3's qwen3_5 model strips them; text generation unaffected).

Reproducible with the scripts in jclyons52/ud2mlx: convert_gguf_to_mlx.py --gguf <Q3_K_XL.gguf> --match-source-quants --hf-path ./Qwen3.8-27B.

Caveats

Text-only: this is the language-model tower of the multimodal checkpoint.
config.json uses model_type: qwen3_5 (loads with mlx-lm 0.31.3+).
MTP/speculative head not included.