Image: reddit.com · MetalQwen3: Full GPU-Accelerated Qwen3 Inference on Apple Silicon with Metal Shaders – Built on qwen3.c, using Claude Code CLI
Qwen3 inference on Apple Silicon GPUs with Metal shaders
A Metal GPU implementation of Qwen3 inference built on qwen3.c, reaching about 75 tokens a second on an M1 Max.
Credits
- Reddit poster QuanstScientist Source ↗
- Repository account BoltzmannEntropy Source ↗
Earliest dated source:
The full story
MetalQwen3 is a Metal compute shader version of Qwen3 transformer inference for macOS, built on Adrian Cable's educational qwen3.c project. The poster says AI coding agents including Claude Code cut the work from months to weeks.
All the core operations (RMSNorm, quantised matrix multiplication, softmax, SwiGLU, RoPE and multi-head attention) run on the GPU with no CPU fallbacks. It reaches around 75 tokens a second on an M1 Max, about 2.1 times the CPU baseline. The poster says it is educational and not meant to replace llama.cpp or vLLM.
How it was done
Tools
Themes
Loading themes…
Sources
From a short post by the maker, so the details are as they tell them.
- Reddit: MetalQwen3: Full GPU-Accelerated Qwen3 Inference on Apple Silicon with Metal Shaders – Built on qwenReddit poster: QuanstScientistPosted
The maker's post on r/ClaudeAI.
If it disappears: archived copy of Reddit: MetalQwen3: Full GPU-Accelerated Qwen3 Inference on Apple Silicon with Metal Shaders – Built on qwen (opens in a new tab)
- The code: github.comRepository account: BoltzmannEntropy
The project on GitHub.