LM/Qwen 3.8 27B

llama.cpp

• https://unsloth.ai/docs/basics/dynamic-3.0-ggufs

Running

# works
./llama.cpp/sycl/bin/llama-server \
  -hf unsloth/Qwen3.8-27B-GGUF:UD-Q6_K_M \
  --spec-type draft-mtp --spec-draft-n-max 2 -ngld 99 \
  -c 64000 -ctk q8_0 -ctv q8_0 \
  -fa on -ngl 99 --load-mode mlock -np 1 --jinja --reasoning-preserve -t 8

# with turboquant
./llama.cpp/vulkan/bin/llama-server \
  -hf unsloth/Qwen3.8-27B-GGUF:UD-Q6_K_M \
  --spec-type draft-mtp --spec-draft-n-max 2 -ngld 99 \
  -c 64000 -ctk q8_0 -ctv turbo4 \
  -fa on -ngl 99 --load-mode mlock -np 1 --jinja --reasoning-preserve -t 8

# assign draft model (cannot run)
./llama.cpp/sycl/bin/llama-server \
  -hf unsloth/Qwen3.8-27B-GGUF:UD-Q6_K_M \
  -hfd unsloth/Qwen3.8-27B-GGUF:MTP/mtp-Qwen3.8-27B-Q4_0.gguf \
  --spec-type draft-mtp --spec-draft-n-max 2 -ngld 99 \
  -c 64000 -ctk q8_0 -ctv q8_0 \
  -fa on -ngl 99 --load-mode mlock -np 1 --jinja --reasoning-preserve -t 8 -lv 5

# turboquant (still not supported in xpu)
./llama.cpp/sycl/bin/llama-server \
  -hf unsloth/Qwen3.8-27B-GGUF:UD-Q6_K_M \
  --spec-type draft-mtp --spec-draft-n-max 4 -ngld 99 \
  -c 64000 -ctk turbo3 -ctv turbo3 \
  -fa on -ngl 99 --load-mode mlock -np 1 --jinja --reasoning-preserve -t 8

Benchmarking

./llama.cpp/sycl/bin/llama-bench \
  -hf unsloth/Qwen3.8-27B-GGUF:UD-Q5_K_XL \
  -ctk q8_0 -ctv q8_0 -fa on -ngl 99 \
  -p 512,1024 -n 512

./llama.cpp/sycl/bin/llama-bench \
  -hf unsloth/Qwen3.8-27B-GGUF:UD-Q6_K_M \
  -ctk q8_0 -ctv q8_0 -fa on -ngl 99 \
  -p 512,1024 -n 512

| model                          |       size |     params | backend    | ngl | type_k | type_v |  fa |            test |                  t/s |
| ------------------------------ | ---------: | ---------: | ---------- | --: | -----: | -----: | --: | --------------: | -------------------: |
| qwen35 27B Q5_K - Medium       |  19.43 GiB |    27.32 B | SYCL       |  99 |   q8_0 |   q8_0 |   1 |           pp512 |        866.11 ± 1.79 |
| qwen35 27B Q5_K - Medium       |  19.43 GiB |    27.32 B | SYCL       |  99 |   q8_0 |   q8_0 |   1 |          pp1024 |        856.36 ± 2.14 |
| qwen35 27B Q5_K - Medium       |  19.43 GiB |    27.32 B | SYCL       |  99 |   q8_0 |   q8_0 |   1 |           tg512 |         20.78 ± 0.05 |

| model                          |       size |     params | backend    | ngl | type_k | type_v |  fa |            test |                  t/s |
| ------------------------------ | ---------: | ---------: | ---------- | --: | -----: | -----: | --: | --------------: | -------------------: |
| qwen35 27B Q6_K                |  21.49 GiB |    27.32 B | SYCL       |  99 |   q8_0 |   q8_0 |   1 |           pp512 |       775.98 ± 15.17 |
| qwen35 27B Q6_K                |  21.49 GiB |    27.32 B | SYCL       |  99 |   q8_0 |   q8_0 |   1 |          pp1024 |        842.36 ± 5.52 |
| qwen35 27B Q6_K                |  21.49 GiB |    27.32 B | SYCL       |  99 |   q8_0 |   q8_0 |   1 |           tg512 |         19.15 ± 0.03 |

Experiment

hf models ls --search "Qwen3.8-27B" --apps llama.cpp --expand "downloads,likes,lastModified" --sort downloads --no-truncate --limit 10

hf models ls --tree -R -h unsloth/Qwen3.8-27B-GGUF

# --mtp --mmproj-auto seem useless
./llama download --mtp --mmproj-auto -hf unsloth/Qwen3.8-27B-GGUF:UD-Q5_K_XL

hf download unsloth/Qwen3.8-27B-GGUF --include "*UD-Q4_K_XL*"
hf download unsloth/Qwen3.8-27B-GGUF --include "*UD-Q5_K_XL*"
hf download unsloth/Qwen3.8-27B-GGUF --include "*UD-Q6_K_M*"
hf download unsloth/Qwen3.8-27B-GGUF --include "mmproj-BF16*"
hf download unsloth/Qwen3.8-27B-GGUF --include "*mtp-Qwen3.8-27B-Q4_0.gguf"

tree -h ~/.cache/huggingface/hub/models--unsloth--Qwen3.8-27B-GGUF

Accuracy

https://unsloth.ai/docs/~gitbook/image?url=https%3A%2F%2F3215535692-files.gitbook.io%2F%7E%2Ffiles%2Fv0%2Fb%2Fgitbook-x-prod.appspot.com%2Fo%2Fspaces%252FxhOjnexMCB3dmuQFQ2Zq%252Fuploads%252Fl9piThTmmUsePst5w3F2%252Fimage.png%3Falt%3Dmedia%26token%3D821e88e2-ad4e-40bf-adf2-e13837228e82&width=768&dpr=3&quality=100&sign=dd5f7cb5&sv=2