Hardware Corner Mar 2026: RTX 4070 Qwen3 14B Q4_K decode @ 16k ctx
32.7 tok/s (decode, per-request, batch 1, KV depth 16k)
Fields
id
benchmarks/hc26-rtx4070-qwen3-14b-q4k-tg-16kMetric
decode_tok_s
Value
32.7
Unit
tok/s (decode, per-request, batch 1, KV depth 16k)
Model
Qwen3 14B Q4_K
Methodology
THIRD-PARTY LAB MEASURED (Hardware Corner per-GPU LLM benchmark page, Updated Benchmarks: March 2026). llama.cpp, single GPU, weights fully in VRAM. Context length is the KV depth the table column names. Token Generation column = batch-1 per-request decode tok/s, NOT aggregate across concurrent requests. Prompt Processing column = prefill tok/s at that context depth. Engine build hash, CUDA version and exact prompt/output token counts beyond the context-depth bucket are not published. Where the page header quotes a single '14B @ 4k' figure that disagrees with the per-model table, the per-cell table value is the one recorded.
Measured by
third_party
Reproducible
yes
Thermal state
unknown
notes
Single RTX 4070 12GB GDDR6X, 504 GB/s as stated on the page. format_id is the nearest standard GGUF in this repo; the page labels the quant Q4_K. Header quotes 32.7 t/s labeled '14B @ 4k'; this 16k cell is 32.7, so the header is this number mislabeled by context depth.
Notes
Single RTX 4070 12GB GDDR6X, 504 GB/s as stated on the page. format_id is the nearest standard GGUF in this repo; the page labels the quant Q4_K. Header quotes 32.7 t/s labeled '14B @ 4k'; this 16k cell is 32.7, so the header is this number mislabeled by context depth.