Sign up

itrejomx/Qwen3.8-27B-Uncensored-HauhauCS-Aggressive-MTPLX-4bit

🤗 Hugging Face sourceapache-2.026.9B params54 GBsafetensors✓ HF checksums availableupdated today
No torrent yet →

Qwen3.8-27B-Uncensored-HauhauCS-Aggressive-MTPLX-4bit

MLX 4-bit build of HauhauCS's Qwen3.8-27B Uncensored Aggressive fine-tune with a working multi-token-prediction (MTP) head, forged and verified with MTPLX (contract=true).

At ~16 GB this fits 24/32 GB Macs. If you have 32 GB+ and want higher fidelity, use the 6-bit build.

Measured speed (M-series Mac, 48 GB)

Mode decode tok/s vs AR
Autoregressive 14.9 1.00x
MTP depth 1 31.0 2.08x
MTP depth 2 37.3 2.51x
MTP depth 3 32.5 2.18x

MTPLX auto-tunes the depth per machine; on the build machine depth 2 was selected.

Usage

mtplx pull <this-repo>
mtplx tune --model <this-repo> --retune   # calibrate MTP depth for your machine

Provenance and deliberate changes

This is a community conversion, not an official HauhauCS release (their releases carry Ed25519-signed manifests; this one does not).

  • Quantization lineage: original bf16 → author's Q8_K_P GGUF → dequantized f16 → MLX 4-bit affine quantization (group size 64). Note this stacks a third quantization layer on an ablated fine-tune; it passed MTPLX Forge's output-quality verification, but for maximum fidelity prefer the 6-bit build. The MTP head is kept in bf16 (mtp_policy: keep_bf16) and was verified against the trunk.
  • Conversion pipeline: llama.cpp's qwen35 GGUF applies a tiled V-head reorder to the linear-attention tensors and stores A_log as -exp(A_log); both transforms were inverted to restore the HF/MLX layout before forging.
  • Chat template: thinking is disabled by default (the original template defaults to enabled). Pass enable_thinking: true to turn it back on.
  • Tokenizer: rebuilt from the family reference tokenizer to restore <think>/</think> special-token registration, which the GGUF metadata round-trip loses.

Credits