Hardware Corner per-GPU LLM pages (Qwen3 8B/14B Q4_K and gpt-oss 20B MXFP4, context sweep March 2026)

Per-GPU pages under /gpu-llm-benchmarks/<sku>/, updated March 2026. Tables give prompt-processing tok/s and token-generation tok/s for Qwen3 8B Q4_K, Qwen3 14B Q4_K, and (16GB cards) gpt-oss 20B MXFP4 at 4k/16k/32k/64k/128k context. Engine is llama.cpp (same lab as the Definitive GPU Ranking page: Ubuntu, CUDA llama-bench, single GPU, batch 1). Per-request, not aggregate. The hero 'Gen (14B 4-bit)' callout on the 4070 Ti SUPER page (84.3 t/s) does not match that page's own 16k grid cell (47.2 t/s) and equals the RTX PRO 5000 Blackwell callout; grid cells are the numbers recorded, hero callouts are not. Q4_K is their label; nearest repo format is gguf-q4-k-m. gpt-oss 20B is MXFP4 (format mxfp4).