GLM-5.3-Flash now runs fully local on the Mac
Rapid-MLX 0.13.3 runs Z.ai's MIT-licensed GLM-5.3-Flash fully local on Apple Silicon; the 4-bit build averages 30 tokens per second on an M3 Ultra.
GLM-5.3-Flash, the model Z.ai published as open weights under an MIT license on August 26, now runs on a Mac without touching the cloud. Rapid-MLX, an inference engine built for Apple Silicon, shipped production-ready support for it in version 0.13.3.
30 tokens per second on an M3 Ultra
The engine, which Raullen Chai built on top of Apple's MLX framework, averages 30 tokens per second with the 4-bit checkpoint across a 512-token generation on an M3 Ultra. On a 192 GB unified-memory machine the model holds 165 GB of active memory, which fits the 320B-total, 18B-active MoE architecture onto a single desktop.
Installation is a single "pip install -U rapid-mlx"; a desktop app is available at rapidmlx.com. Because the engine exposes an OpenAI-compatible API, tools like Claude Code and Cursor can point at the local model directly.
A chart shared with the announcement places GLM-5.3-Flash at 57 on the Intelligence Index, just below the Claude Opus family and above most large models. Z.ai covered the 1M-token context window and the MIT license in its launch post.
