Two Days, 57 Runs: Getting a 27B Local Model from 10 to 105 tok/s
57 runs, one 27B model, two days of server settings: 10 tok/s to 105, and most of the gain came from a single KV cache flag.
// latest from the lab
Fifty copy-paste prompts in five buckets — thinking, storytelling, philosophy, something weird and self-improvement. Paste one, see what comes back.
57 runs, one 27B model, two days of server settings: 10 tok/s to 105, and most of the gain came from a single KV cache flag.