A 54GB model in one 8GB file, with agent skills intact. Underdog Saluki 27B 1.0 (Conway Research, October 8) rebuilds Qwen3.8-27B as a tuned 2-bit mix: one 7.89GB GGUF that stock llama.cpp serves, with tool calling scored at 88 against the full model’s 84 on the team’s own bench, parallel tool calls at 42 against 35, and 96 percent average knowledge retention across 9 tests.
The trade is tuned for agents, not leaderboards. Most 2-bit quants chase benchmark numbers and quietly lose the function-call format that turns a chat model into an agent. Saluki’s compression is built the other way around: the agentic behaviors are what it is tuned to keep, and the vision cases are covered by an optional mmproj file (928MB at F16) rather than baked into the main file.
Who it serves. An 8GB machine, a plain mini PC, any card from the last decade: the whole 27B class runs without a rental. Apache 2.0, thinking mode on by default and switchable per request.