DEPT·F1 — TOOLING & PLATFORM2026
ggufbench
An agentic benchmark for local GGUF models: prompt speed, tool-calling, code generation and multi-step agent work graded under one identical harness, with the raw receipts of every call — 100% local, no cloud, no APIs.
The brief
Public benchmarks are trained on, and “how smart is it” is not the question a local rig answers. ggufbench asks which GGUF to run on your own machine: every model runs the same harness, every scoreboard row states the task set it really ran, and exclusions are recorded with a reason — never silent.
What was built
Four dimensions under one identical harness: prompt speed, tool-calling, code generation and multi-step agent work — three runs per task (seeds 42/43/44) on the same 2× RTX 5080 rig, reported as mean ± stdev.
The current task set: 18 tool, 72 code and 72 agent tasks — versioned (v1 → v5) so every scoreboard row states the set it really ran; a single 32768-token reasoning budget and thinking modes measured separately.
100% local — no cloud, no APIs, no telemetry — and receipts included: every call leaves timings, tool calls and per-task traces for a live scoreboard, and 13 retired models stay as record only.