Hardware Corner per-GPU local LLM benchmark pages (llama.cpp Qwen3 Q4_K and gpt-oss-20B MXFP4 context sweep, March 2026)

Index of 21 in-lab GPU pages. Each page publishes prompt-processing and token-generation tok/s at context tiers 4k/16k/32k/64k/128k/256k for Qwen3 8B and 14B Q4_K GGUF, and on 16GB+ cards gpt-oss 20B MXFP4. Pages say Updated Benchmarks March 2026 and describe llama.cpp llama-bench, but do not print the commit, prompt length, generation length, or flash-attention flag. Distinct from hardware-corner-gpu-ranking-local-llm (older Q4_K_XL ranking, 16k decode column already ingested) and from hardware-corner-rtx-5090-llm-benchmarks (different URL, Q4_K_XL sweep). Per-request batch-1, not aggregate. Third-party measured.