Hardware Corner Mar 2026: RTX 4060 Ti 16GB gpt-oss 20B MXFP4 prefill @ 16k ctx

THIRD-PARTY LAB MEASURED (Hardware Corner per-GPU LLM benchmark page, Updated Benchmarks: March 2026). llama.cpp, single GPU, weights fully in VRAM. 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. Engine build hash, CUDA version and exact prompt/output token counts beyond the context-depth bucket are not published. Where the page header quotes a single '14B @ 4k' figure that disagrees with the per-model table, the per-cell table value is the one recorded.

Single RTX 4060 Ti 16GB GDDR6, 288 GB/s as stated on the page. format_id is the MXFP4-MoE GGUF this repo already tracks; the page labels the quant MXFP4. 16 GB is what makes the 128k column possible for a 20B MoE.