license: apache-2.0
base_model:

• Qwen/Qwen3.8-27B
• orcarouter/Qwen3.8-27B-Uncensored
  base_model_relation: quantized
  library_name: mlx
  pipeline_tag: image-text-to-text
  tags:
• mlx
• 4-bit
• mtp
• speculative-decoding
• abliterated
• uncensored
• qwen3.8
  language:
• en
• zh

Qwen3.8-27B-Uncensored-MLX-4bit-MTP

Uncensored Qwen3.8-27B (orcarouter abliteration lineage), MLX uniform 4-bit,
with the native MTP speculative-decoding head grafted back in.

Provenance

• Body: orcarouter/Qwen3.8-27B-Uncensored-MLX (root files = 4-bit).
  Shard bytes are unchanged (hardlink-verified); only filenames/index were
  rewritten from -of-00003 to -of-00004.
• MTP head: orcarouter/Qwen3.8-27B-Uncensored-MLX mtp/model.safetensors
  sidecar (15 BF16 tensors), re-indexed as language_model.mtp.* in
  model-00004-of-00004.safetensors - the embedded layout oMLX native MTP
  expects (same as Jundot/Qwen3.8-27B-oQ4e-mtp).
• tokenizer.json / vocab.json: byte-verified copies from
  mlx-community/Qwen3.8-27B-4bit (identical LFS objects to the orcarouter repo).
• Abliteration: 129 residual-writer tensors edited vs stock MLX 4-bit
  (dequantized Frobenius verification: 16 self_attn.o_proj + 48
  linear_attn.out_proj + 64 mlp.down_proj + embed_tokens; norms and vision
  tower bit-identical to stock).

Verification

• Index <-> shard headers consistent (2195 tensors).
• SHA256SUMS.json: full sha256 of the new MTP shard; body shards referenced
  by their upstream LFS sha256.

Use

oMLX: add to a scanned model dir, load, check engine log for
Lightning MTP enabled ... (checkpoint layout: mtp.). Depth 1 (prose) or 2
(code); depth 3 gains nothing on this architecture.

Requires oMLX with indexed-sidecar/MTP detection (>=0.6.x; issues #1944/#1962/#2220 fixed).

License: Apache-2.0, inherited from Qwen/Qwen3.8-27B.