MLX Benchmark

SkillAI & models

Run and compare MLX Vulkan inference benchmarks. Use when asked to benchmark this repository, measure a change, run bf16 or 8-bit MLX performance tests, or compare current results with the README baseline.

Available today. Use it from your connected AI after setup.

Connect ahel once, and every AI you use reads what you have installed.

Then ask your AI: use the MLX Benchmark skill

What this skill tells your AI

The instructions your AI receives, as published by goniz/mlx-vulkan in .agents/skills/mlx-benchmark/SKILL.md and read by ahel’s review.

Run benchmarks through ./dev.sh only. Treat any user-supplied model or benchmark option as an override; otherwise run both standard precisions.

  1. Read README.md and identify the relevant current baseline numbers.
  2. Run ./dev.sh benchmark bf16.
  3. After it completes, run ./dev.sh benchmark 8bit.
  4. Never run benchmark commands concurrently or in parallel tool calls. ./dev.sh benchmark already serializes via /tmp/mlx-gpu/gpu.lock (shared with CI); still launch them one after another so waiters are ordered intentionally.
  5. Compare each result with the comparable README baseline. State the absolute and percentage difference when the metrics permit it; otherwise explain the direct comparison.

Report the commands run, results for each precision, the baseline used, and the performance difference. Do not change code unless explicitly asked.

Signals

GitHub stars
38
Forks
5
Last commit
Aug 2026
Advanced
Catalog kind
skill
Gateway key
mlx-benchmark
Source
github.com/goniz/mlx-vulkan