DeepSeek-R1-0528-Qwen3-8B · llama.cpp · Q4_K_M

status: done

Configuration

Modeldeepseek-ai/DeepSeek-R1-0528-Qwen3-8B
CompanyDeepSeek
FamilyDeepSeek
Parameters8B (dense)
Enginellama.cpp
Quant / precisionQ4_K_M
Why this quantGGUF Q4_K_M from unsloth (trusted quantizer) — widest llama.cpp coverage, strong size/quality balance.
Downloadunsloth/DeepSeek-R1-0528-Qwen3-8B-GGUF
Context window65536
Input modalitiestext

Measured results

Prefill tok/s256.08
Decode tok/s374.69
Peak memory (GB)24.22 (system MemAvailable delta (10s sampling))
Completed2026-06-22 04:10 +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-R1-0528-Qwen3-8B-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-R1-0528-Qwen3-8B-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-r1-0528-qwen3-8b DeepSeek Q4_K_M 5-15B conc-32

Notes

A reasoning distill on a Qwen3-8B base — the fastest-decoding dense 8B in the set. DeepSeek’s R1-0528 reasoning trace distilled onto Qwen3-8B, Q4_K_M from unsloth.

  • Workload: ShareGPT V3, concurrency 32. 982/1000, 18 errors (slot-split) in 667 sdid not hit the time cap. Loaded in ~17 s.
  • Throughput (aggregate, conc 32): prefill 256.1 tok/s, decode 374.7 tok/s — edging past Llama-3.1-8B (365) and clear of the Granite-4.1-8B hybrid (324) on the identical engine/quant/ctx. TTFT median 745 ms, TPOT median 70.7 ms (≈14 tok/s/stream), req throughput 1.47/s.
  • It’s a reasoning model, and it shows in the token mix. 249.7 k completion tokens over 982 requests ≈ 254 each — essentially every request ran to the 256 max-tokens cap (the model keeps thinking), vs ~215 each for Llama. That long-output behaviour, plus the heavier prefill batching, drives the high TTFT (745 ms) — first tokens wait behind fuller decode batches — even though steady-state per-token decode (70.7 ms TPOT) is the best of the dense 8Bs. Prefill tok/s (256) reads low for the same reason: the run is decode-dominated, so less of the wall-clock is spent on prefill.
  • Memory: 24.2 GB — between Llama-3.1-8B (22.7) and Granite-8B (25.4), as expected for a dense Qwen3-8B with the standard full-64K transformer KV pre-allocated.
  • Slot-split errors (18): -c 65536 --parallel 32 → 2048 tok/slot; longer ShareGPT prompts 400. Engine-config artifact, consistent across all llama.cpp runs.