Hardware Corner: gpt-oss-20B on an RTX 3090, Flash Attention context sweep

Allan Witt, Hardware Corner, last updated 2025-10-26. Hands-on llama.cpp / LM Studio sweep of lmstudio-community gpt-oss-20b-MXFP4.gguf on a single 24 GB RTX 3090, full GPU offload. Three tables: Windows 11 LM Studio (context ceiling 36k), Linux llama.cpp server without Flash Attention (ceiling 50k), and a 2025-10-26 Linux re-run with Flash Attention that reaches 131,000 tokens. The FA table publishes prompt-processing and token-generation tok/s at 4k, 8k, 16k, 32k, 45k, 57k, 65k, 86k and 131k. The page does not name the llama.cpp build, CUDA version, flash-attention flag, batch size, or the exact prompt and generation lengths behind each cell. Third-party measurement, not a vendor claim.