Qwen3.8-27B Coder390 EfficientThink — Q2 LynnStyle MTP P1M
P1M vocabulary-pruned version of the Q2 LynnStyle MTP GGUF from
nerkyor/Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-MTP-DFlash2, downloaded at revision 21418f49094bca7df56df0f7c3e94d33057831a1.
P1M retains ASCII, common math and typography symbols, special tokens, all 256 byte-fallback tokens, and partial UTF-8 fragments. Vocabulary size was reduced from 248,320 to 129,272. The vocabulary matrices were gathered directly in quantized space without dequantizing or requantizing; all other tensors, including the embedded MTP block, are byte-identical to the source. Non-ASCII text outside the P1M keep-set remains representable through byte fallback, with a higher token cost.
File
| File | Quantization | Size |
|---|---|---|
Qwen3.8-27B-Coder390-EfficientThink-Q2LynnStyle-MTP-P1M.gguf | Q2 LynnStyle with embedded MTP | 12.43 GB |
Use with llama.cpp
llama-cli -m Qwen3.8-27B-Coder390-EfficientThink-Q2LynnStyle-MTP-P1M.gguf -p "Write a short Python function to sort a list."
The upstream card describes its Q2 GGUF as containing an MTP draft head in blk.64; no external draft model is required. See the upstream model card for training, evaluations, and full attribution. The license shown here follows the upstream repository's Apache-2.0 declaration.