benchmarks / qwen3-6-35b-a3b--dgx-spark-x1--llama-cpp--q4-k-m--dflash-draft--p2048-o256-c1--2026-07-06

Qwen3.6 35B-A3B on NVIDIA DGX Spark

compare

81.8tok/s

1.18sttft

1.9kprefill

llama.cpp b15-ee445f9Q4_K_Mspec-decodingJul 6, 2026

Run configuration

nodes

batch size

—

concurrency

context

10240

prompt / output

2042 / 256

weights unsloth/Qwen3.6-35B-A3B-GGUF ↗draft model z-lab/Qwen3.6-35B-A3B-DFlash ↗

Metrics

decode tok/s

81.8

prefill tok/s

1.9k

ttft

1.18s

total tok/s

59.2

inter-token

12ms

e2e latency

4.30s

req/s

0.231

peak memory

32.9 GB

avg power

47.2 W

energy/req

204.2 J

temperature

draftAcceptanceRate

0.43

Reproduce

~/Dev/llama.cpp/build/bin/llama-server -m ~/models/gguf/Qwen3.6-35B-A3B-UD-Q4_K_M.gguf -c 10240 --parallel 1 -ngl 99 -fa on --host 127.0.0.1 --port 8080 -md ~/models/gguf/Qwen3.6-35B-A3B-DFlash-F16.gguf --spec-type draft-dflash --spec-draft-n-max 5
python3 bench/harness.py bench/configs/qwen3-6-35b-a3b--dflash-draft--llama-cpp.json --rows-out data/benchmarks/qwen3-6-35b-a3b.json

spec-decoding profile: greedy with z-lab DFlash drafter (0.39B, trained on original Qwen3.6, converted to GGUF), draft-dflash spec-max 5, original target. For the heretic target this measures fine-tune drift on draft acceptance vs the matched original.