RTX 5070 llama.cpp: Qwen3 8B Q4_K decode at 4k ctx (Hardware Corner)
85.8 tok/s token generation, batch 1, per-request, ctx 4k
Fields
id
benchmarks/hc-5070-qwen3-8b-q4k-4k-tgMetric
decode_tok_s
Value
85.8
Unit
tok/s token generation, batch 1, per-request, ctx 4k
Model
Qwen3 8B
Methodology
Hardware Corner per-GPU context-tier table (March 2026 update), single RTX 5070, llama.cpp GGUF, batch-1 per-request. Qwen3 8B Q4_K, context depth 4k. Prefill column is prompt-processing tok/s; generation column is decode tok/s. NOT an aggregate across concurrent requests. The page does not reprint the llama-bench build hash (the site's sibling ranking page pins CUDA 12.8 / Q4_K_XL).
Measured by
third_party
Reproducible
yes
notes
Paired prefill is 3487.7 tok/s in the same table cell. Per-request, not aggregate.
Notes
Paired prefill is 3487.7 tok/s in the same table cell. Per-request, not aggregate.