benchmarks / qwen3-5-27b--dgx-spark-x1--llama-cpp--q4-k-m--p2048-o256-c1--2026-07-06
Qwen3.5 27B on NVIDIA DGX Spark
compare12.0tok/s
2.72sttft
790prefill
llama.cpp b1-2da6686Q4_K_MJul 6, 2026
Run configuration
nodes
1
batch size
—
concurrency
1
context
10240
prompt / output
2040 / 256
weights unsloth/Qwen3.5-27B-GGUF ↗
Metrics
decode tok/s
12.0
prefill tok/s
790
ttft
2.72s
total tok/s
10.6
inter-token
83ms
e2e latency
24.06s
req/s
0.042
peak memory
28.1 GB
avg power
43.3 W
energy/req
1042.3 J
Reproduce
~/Dev/llama.cpp/build/bin/llama-server -m ~/models/gguf/Qwen3.5-27B-Q4_K_M.gguf -c 40960 --parallel 4 -ngl 99 -fa on --host 127.0.0.1 --port 8080 python3 bench/harness.py bench/configs/qwen3-5-27b--llama-cpp--q4-k-m.json --rows-out data/benchmarks/qwen3-5-27b.json
unsloth GGUF, llama.cpp CUDA build for GB10 (sm121), flash attention on, 4 slots. Dense counterpart to Qwen3.5-35B-A3B MoE.