Lightning-MLX: Fastest Local AI Engine for Apple Silicon
Benchmark the engine on your own workloads to validate performance claims.
Benchmark the engine on your own workloads to validate performance claims.
Summary
Lightning-MLX is a new local AI engine built for Apple Silicon, designed specifically for agentic workflows such as coding agents, tool calling, and short‑turn interactions.
The project, hosted on GitHub at github.com/samuelfaj/lightning-mlx, delivers impressive throughput on a MacBook Max M5 with 128 GB RAM: Qwen3.6‑27B achieves 40.67 tokens per second, while Qwen3.6‑35B‑A3B reaches 220.86 tokens per second. The author invites feedback on benchmark designs, the MTPLX preset defaults, and additional Apple Silicon configurations to further validate performance.
Overall, Lightning‑MLX claims to be the fastest local engine for Apple Silicon, offering a high‑performance, agent‑friendly foundation for developers looking to run large language models natively on macOS.
Key changes
- Fastest local AI engine for Apple Silicon
- Optimized for agentic use focusing on coding agents, tool calling, and short‑turn workflows
- Supports Qwen3.6‑27B at 40.67 tok/s
- Supports Qwen3.6‑35B‑A3B at 220.86 tok/s
- Repo at github.com/samuelfaj/lightning-mlx
- Feedback requested on benchmark designs, MTPLX preset defaults, and other Apple Silicon setups