ComputingForGeeks: Run Qwen3.8-27B locally, VRAM and 262K context measured on an RTX 4090

Third-party measurement, read 2026-10-18. The author says every number was measured in September 2026 on an RTX 4090 (Ollama 0.34.4, llama.cpp build b11200) and an RTX PRO 6000 Blackwell Server Edition (vLLM 0.30.0), both Ubuntu 24.04. llama.cpp table on the 4090, UD-Q4_K_XL, MTP off: 64K/f16 uses 20.6 GiB and decodes at 47.0 tok/s; 131K/q8_0 uses 21.3 GiB at 46.6 tok/s; 262K/q4_0 is the only combination that loads, at 22.2 GiB and 46.5 tok/s (262K/q8_0 does not load). Going from 8K to 64K added 3,640 MiB for 57,344 tokens, which the post reads as about 64 KiB per token at f16. Needle test: a 196,971-token log with the passphrase buried 60 percent of the way in, thinking off, answered TANGERINE-4471. Prefill 121.5 s (1,621 tok/s), generation at that depth 26.6 tok/s over a 9-token answer, VRAM 22.2 GiB right after the request. Server flags printed: llama-server -ngl 99 -fa on -c 262144 -ctk q4_0 -ctv q4_0. Raw llama-bench generation without MTP: 47.9 tok/s Q4_K_XL, 41.5 Q5_K_XL, 39.6 Q6_K, prompt processing 2,600-3,000 tok/s (short context, not the 197K depth). The vLLM matrix on the 96 GB PRO 6000 (1,024 in / 512 out) is a different machine and is not the number this record uses. No raw JSON is linked, so confidence stays at a blog measurement.