Rapid-MLXFrom the builder

Rapid-MLX 0.11.0 — 25.6× faster to first token, meet your Mac's new agent

Rapid-MLX, an open-source local inference engine for Apple Silicon, released version 0.11.0 and says reusing a 6,000-token prefix on an M3 Ultra cut time to first token from 13.1 seconds to 0.51, while an opt-in response cache returns identical requests in 2 milliseconds instead of 656. Its rapid-mlx chat command is now an agent that chains tools from any MCP server and supports thinking budgets. The release adds five model families including Tencent Hunyuan 3 and constrains tool-call decoding at the grammar level so even reasoning models cannot emit malformed calls. The project also lists open gaps, such as parked native MTP for Qwen3.6 and no same-machine comparison with llama.cpp or Ollama this cycle.

Official blogJul 24, 20264 min read