Qwen3.8-27B-MTP-q27

Qwen3.8-27B-MTP q27 — c-small

This is the reproducible c-small q27 pack selected for the q27 Metal v0.7.0 24 GB release. It is a text-only/agentic quantized derivative of Qwen/Qwen3.8-27B. The vision tower, mmproj, multimodal MRoPE, and image/audio/video input are not included or supported.

Immutable provenance

• Official checkpoint: Qwen/Qwen3.8-27B@1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0
• Conversion source: ggml-org/Qwen3.8-27B-GGUF@0669b98607d47046c7c2b3f801011d54a08cfccf
• Base BF16 GGUF: 53,808,281,952 bytes, SHA-256 5a3eedc837bcbd1365cdbf5b71e698df3122e76586ba872f07ce3ed4a9bfa97e
• MTP BF16 GGUF: 5,946,009,888 bytes, SHA-256 5723e551c4ee2b8c11933d687f445800521047edc39769bee6e3c78c4e141ace
• q27 artifact: 15,704,702,464 bytes, MD5 e43dc81606d28825bfe129b56401220d, SHA-256 e7736cbfabccc61a57b19468217b8d1975c589c56bd15fdda121fdcc0e3ae0ff
• tokenizer: 7,212,628 bytes, SHA-256 f8adac2001668f3ba68e040e5c13f0982d6a194e2b4d9047cc050fdc6d7c93f2

CHECKSUMS.txt records the complete reproduction manifest.

Recipe

Using q27's hash-locked repack environment:

python tools/repack.py Qwen3.8-27B-BF16.gguf qwen38-27b-mtp-c-small.q27 \
  --mtp mtp-Qwen3.8-27B-BF16.gguf \
  --q4-head \
  --q8 '^blk\.[0-9]+\.attn_output\.weight$' \
  --tag q38-c-small-v1
python tools/export_tokenizer.py Qwen3.8-27B-BF16.gguf qwen38-27b-mtp.tok

The policy keeps the shared language-model head at Q4 and promotes each attn_output matrix to Q8. It intentionally does not promote ssm_out. See docs/metal/QWEN38_REPACK.md in q27 for the split-GGUF validation and hash-locked dependency procedure.

Selection evidence

On the exact 24 GiB Apple M4 release machine, this artifact preserved a 34,714-token one-slot FP16 auto-context envelope (q4s: 36,409), improved single-pass 24,576-token WikiText-2-raw PPL from 5.247 to 5.233 in every reported position band, and passed the registered Metal canonical, Chat, Responses, Anthropic, and native-agent tool gates. Plain 64-token decode was 3.33 tok/s versus q4s 3.38 tok/s; 128-token MTP decode was 3.63 versus 3.73 tok/s. The measured quality gain was selected over the small footprint and throughput cost.

License and attribution

The upstream model is Apache-2.0. This repository includes the upstream license and an attribution notice. Use of this derivative remains subject to the Apache License, Version 2.0.