A native Zig AI server promises faster model runs on Mac
A new written in Zig aims to fix slow support for newly released AI models on Mac, along with unreliable tool use, image understanding, and context handling. It runs natively without Python or Electron and supports most MLX and GGUF models.
It accepts the same API formats used by OpenAI, Anthropic, and Ollama, making it easier to connect existing software. In the developer's published tests across Gemma 4 and Qwen 3.6, it was 48% faster than LM Studio by geometric mean; the includes one script for reproducing those tests.
The server recognizes several model-specific tool-call formats and repairs malformed output before passing it to other software. It also combines local image, video, music, voice-cloning, and in one MLX-based media system.
Key points
- The server is written in Zig and needs neither Python nor Electron.
- It supports most MLX and GGUF models plus OpenAI, Anthropic, and Ollama API formats.
- Published tests report a 48% geometric-mean speed advantage over LM Studio on Gemma 4 and Qwen 3.6.
- It detects different tool-call formats and repairs malformed model output.
- It combines image, video, music, voice-cloning, and in one local system.