Qwen3.8-27B-test-GGUF

Qwen3.8-27B-test GGUF

IQ3_S, Q4_K_M, and Q5_K_M conversions of vvsotnikov/Qwen3.8-27B-test, pinned to revision 13171b198fbeb3a2ecc5aaabe0d3ffd2fc8a39fa.

The source is a 50/50 linear merge of Qwen3.6-27B and Qwen3.8-27B. Original merge provenance is in merge-manifest.json.

The MTP head is embedded in each main GGUF. A separate drafter file is not required. The vision GGUF is shared by all three quantizations and is required for image input.

Run

Use a recent llama.cpp with Qwen3.5 native MTP support. The tested revision is 64e9bceb2c3a856efed96feda784a50947049feb.

llama-server \
  --model Qwen3.8-27B-test-Q4_K_M.gguf \
  --mmproj mmproj-Qwen3.8-27B-test-BF16.gguf \
  --alias Qwen3.8-27B-test \
  --spec-type draft-mtp --spec-draft-n-max 3 \
  --n-gpu-layers 99 --ctx-size 8192 --parallel 2 \
  --jinja --reasoning on --reasoning-format deepseek \
  --chat-template-kwargs '{"enable_thinking":true,"preserve_thinking":true}' \
  --temp 1.0 --top-p 0.95 --top-k 20 --min-p 0 --repeat-penalty 1 \
  --host 127.0.0.1 --port 8080

Substitute the IQ3_S or Q5_K_M filename to use that quant. API requests should retain the source sampling settings: temperature 1.0, top-p 0.95, top-k 20, with thinking enabled.

Conversion

All three quants were produced directly from one BF16 GGUF intermediate, using llama.cpp's default quantization recipes, without an importance matrix. They were not converted from previously quantized MLX weights.

python convert_hf_to_gguf.py /path/to/pinned-bf16-source \
  --outtype bf16 --outfile model-BF16.gguf
./build/bin/llama-quantize model-BF16.gguf model-IQ3_S.gguf IQ3_S 10
./build/bin/llama-quantize model-BF16.gguf model-Q4_K_M.gguf Q4_K_M 10
./build/bin/llama-quantize model-BF16.gguf model-Q5_K_M.gguf Q5_K_M 10
python convert_hf_to_gguf.py /path/to/pinned-bf16-source \
  --mmproj --outtype bf16 --outfile mmproj-BF16.gguf

Environment: Python 3.12.14, PyTorch 2.11.0, Transformers 5.14.0, NumPy 1.26.4; llama.cpp built with Metal on an Apple M5 Max with 128 GiB RAM. Full conversion and validation settings are in conversion-manifest.json.

Validation

All three quants retain all 866 main-file tensors, including all 15 MTP tensors, with one next-token-prediction layer declared in the GGUF metadata.

All three passed local text, image, streaming, two-request concurrency, invalid-request handling, client-disconnect cleanup, and post-cancellation recovery checks on Metal. The image test correctly identified a red square and blue circle. Text responses included separate reasoning content with thinking enabled.

On the first arithmetic smoke request, Q4_K_M accepted 98/114 drafted tokens (86.0%) and Q5_K_M accepted 78/93 (83.9%). These are individual smoke results, not comparative quality or speed benchmarks. No BF16 parity or broad quality evaluation is claimed.

IQ3_S was added on 2026-09-15 using the same BF16 intermediate and llama.cpp revision. It passed the same functional checks with thinking and native MTP enabled. Its arithmetic smoke request accepted 113/144 drafted tokens (78.5%). This is a smoke-test observation, not a quality or performance benchmark. The exact artifact SHA256, input SHA256, and validation results are recorded in conversion-manifest.json.