Hardware Corner RTX 4070 local LLM benchmark page (March 2026)

Per-SKU llama.cpp table, page header 'Updated Benchmarks: March 2026'. Single RTX 4070 (12 GB GDDR6X, 504 GB/s stated). Tables: Qwen3 8B Q4_K at 4k/16k/32k (PP 3,564.1 / 2,064.3 / 1,116.7; TG 71.2 / 52.1 / 38.1) and Qwen3 14B Q4_K at 4k/16k (PP 2,099.9 / 1,355.8; TG 42.5 / 32.7). Header '32.7 t/s generation and 1,356 t/s prompt processing, 14B @ 4k' matches the 16k cells (32.7 and 1,355.8), so the header mislabels the context depth; the per-cell table is what was recorded. No llama.cpp build hash, CUDA version, or exact prompt/output token counts beyond the context-depth bucket.