Hardware Corner Mar 2026: RTX 4090 Qwen3 14B (Q4_K) decode @ 32k ctx
55.4 tok/s (decode, per-request, batch 1, KV depth 32k)
Fields
id
benchmarks/hc26-4090-qwen3-14b-tg-32kMetric
decode_tok_s
Value
55.4
Unit
tok/s (decode, per-request, batch 1, KV depth 32k)
Model
Qwen3 14B (Q4_K)
Methodology
THIRD-PARTY LAB MEASURED (Hardware Corner per-GPU LLM benchmark page, Updated Benchmarks: March 2026). llama.cpp, single RTX 4090 (24 GB GDDR6X). 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. The page does not publish the llama.cpp build hash, CUDA version, or exact prompt and output token counts beyond the context-depth bucket.
Measured by
third_party
Reproducible
yes
Thermal state
unknown
notes
format_id is the nearest standard GGUF in this repo; the page labels the quant as shown in the model field. Single card, weights fully in VRAM. A dash in the source table means the context did not fit and is not recorded.
Notes
format_id is the nearest standard GGUF in this repo; the page labels the quant as shown in the model field. Single card, weights fully in VRAM. A dash in the source table means the context did not fit and is not recorded.