DeepSeek-R1-Distill-Qwen-14B · llama.cpp · Q4_K_M

status: done

Configuration

Modeldeepseek-ai/DeepSeek-R1-Distill-Qwen-14B
CompanyDeepSeek
FamilyDeepSeek
Parameters14B (dense)
Enginellama.cpp
Quant / precisionQ4_K_M
Why this quantGGUF Q4_K_M from bartowski (trusted quantizer) — widest llama.cpp coverage, strong size/quality balance.
Downloadbartowski/DeepSeek-R1-Distill-Qwen-14B-GGUF
Context window65536
Input modalitiestext

Measured results

Prefill tok/s178.42
Decode tok/s243.75
Peak memory (GB)30.85 (system MemAvailable delta (10s sampling))
Completed2026-06-22 04:43 +08

Full run command

# ghcr.io/ggml-org/llama.cpp:full-cuda (dispatcher → --server). GGUF under /home/gauravmm/models.
docker run --gpus all -p 8081:8081 -v /home/gauravmm/models:/models:ro \
  ghcr.io/ggml-org/llama.cpp:full-cuda \
  --server -m /models/DeepSeek-R1-Distill-Qwen-14B-Q4_K_M.gguf -ngl 99 -c 65536 \
  --parallel 32 -cb --host 0.0.0.0 --port 8081
python3 scripts/bench-serving.py --base-url http://localhost:8081 \
  --model DeepSeek-R1-Distill-Qwen-14B-Q4_K_M.gguf \
  --dataset benchmark_data/ShareGPT_V3_unfiltered_cleaned_split.json \
  --num-prompts 1000 --max-seconds 900 --concurrency 32 --max-tokens 256

deepseek-r1-distill-qwen-14b DeepSeek Q4_K_M 5-15B conc-32

Notes

The older R1 distill on a Qwen2.5-14B base — clean dense scaling, and another point against the Gemma KV cliff. Q4_K_M from bartowski.

  • Workload: ShareGPT V3, concurrency 32. 982/1000, 18 errors (slot-split) — completed all dispatched prompts; the run drained at 957 s (just over the 900 s dispatch deadline as in-flight requests finished, so hit_time_cap=false).
  • Throughput (aggregate, conc 32): prefill 178.4 tok/s, decode 243.8 tok/s. TTFT median 711 ms, TPOT median 111.5 ms (≈9 tok/s/stream), req throughput 1.03/s. Sits right where a dense 14B should: below the 8Bs (365–375) and above the compute-heavy Gemma-12B (195).
  • Memory: 30.9 GB — and that’s the headline comparison. A 14B Qwen here uses less memory than the 12B Gemma (41.3 GB). Same quant, same 64K ctx, same engine — the only difference is attention architecture. Qwen2.5’s standard GQA KV is far cheaper than Gemma’s wide global-attention KV, so the larger model has the smaller footprint. Two independent dense data points (this + Llama-3.1-8B) now confirm the Gemma-12B number is an architecture cliff, not a measurement fluke.
  • Reasoning model: 233 k completion tokens over 982 reqs ≈ 237 each — long outputs near the 256 cap, consistent with the R1 distill’s think-then-answer behaviour.
  • Slot-split errors (18): -c 65536 --parallel 32 → 2048 tok/slot; longer ShareGPT prompts 400. Engine-config artifact, consistent across all llama.cpp runs.