---

license: apache-2.0

base_model:

• Qwen/Qwen3.8-27B

pipeline_tag: text-generation

tags:

• gguf
• qwen
• qwen3_5
• unsloth
• imatrix
• conversational

---

Qwen3.8-27B-MTP-GGUF-SPLIT

Split GGUF shards of unsloth/Qwen3.8-27B-GGUF, using the current Unsloth Dynamic v3 UD-Q3_K_XL quantization.

The original quantized tensors are preserved; this is a split, not a requantization.

Included quantization

| Quantization | Total size | Shards | llama.cpp type |

|---|---:|---:|---|

| Qwen3.8-27B-UD-Q3_K_XL | 13,146,393,824 bytes | 3 | Q3_K - Large |

| File | Bytes | SHA-256 |

|---|---:|---|

| Qwen3.8-27B-UD-Q3_K_XL-00001-of-00003.gguf | 4,906,066,688 | 7503289f9bf53bd41e5c29bbf6a125eb5744b3ca5842b8224f5763dc07dda3db |

| Qwen3.8-27B-UD-Q3_K_XL-00002-of-00003.gguf | 4,890,964,352 | 32a142816d057e24e3bc6a55e6523a081559d3828638805779eeca822d8b0c6a |

| Qwen3.8-27B-UD-Q3_K_XL-00003-of-00003.gguf | 3,349,362,784 | 7112823753700a9dfb15289fbbe0006f7c1b301b409cedc0dd9f61549c061291 |

Run with llama.cpp

Download all three shards into the same directory:

hf download Nielk38/Qwen3.8-27B-MTP-GGUF-SPLIT \
  --include "*.gguf" \
  --local-dir ./Qwen3.8-27B-MTP-GGUF-SPLIT

Load only the first shard; llama.cpp discovers the other two automatically:

llama-cli \
  -m ./Qwen3.8-27B-MTP-GGUF-SPLIT/Qwen3.8-27B-UD-Q3_K_XL-00001-of-00003.gguf \
  -ngl 999 \
  -c 8192

The Qwen3.8 checkpoint is MTP-trained. This repository contains the main Q3 GGUF only. The optional standalone MTP draft model and vision projectors can be downloaded from the source GGUF repository.

Provenance and validation

• Source file: Qwen3.8-27B-UD-Q3_K_XL.gguf
• Source revision: 27af057ecb382ddfea5d12837360a8980560e3ed
• Source SHA-256: 8c2a45ff85e7674ca185ec8eb6cdeab0e617ed9d8018caed0b64380eb2a67a5e
• Splitter: llama-gguf-split from llama.cpp b10502 (0adcc3bb5)
• Split command:

llama-gguf-split --split --split-max-size 4900M \
  Qwen3.8-27B-UD-Q3_K_XL.gguf \
  Qwen3.8-27B-UD-Q3_K_XL

Validated locally on 2026-08-19 with llama.cpp b10502 on an AMD Radeon RX 7900 XTX by loading shard 00001 directly. The prompt What is 2+2? Answer with only the numeral. returned 4.

The shards were also merged back with llama-gguf-split --merge. All 866 tensor descriptors and tensor bytes matched the source, with logical tensor SHA-256 2b7efe87386ffa39812f47aac97d431fe1f7706a8733b8cff1833e324e0371f2. The merged container retains llama.cpp's three split.* metadata keys, so its whole-file hash is intentionally different from the unsplit source container.

See SHA256SUMS for machine-readable shard checksums.