davidtai/turboexl3 is a Zig-based Metal kernel library that handles EXL3 quantized weights directly on Apple GPUs. MLX runs the Metal kernels, and no separate release version has been published. According to the repository description, it restores weights compressed with trellis code inside the kernel.
This design sends the routing experts of a mixture of experts (MoE) layer through a split matrix-vector operation path, while attention, the gated delta net, shared experts and the output layer go through a matrix multiplication path. Weights are not unpacked into a separate copy first but are wired into the computation directly. The repository includes tests that compare against a host reference implementation: substituted kernels are judged on byte-level identity, and kernels whose summation order differs are given a tolerance. It also ships a performance and coding evaluation document. The project runs as an external quantization plugin for mlx-serve, connecting it to a local inference path on Apple Silicon.
That said, there is no public latest release, and the license information is shown as NOASSERTION, so it cannot be confirmed. The presence of an evaluation document alone does not establish the real performance or an advantage over other inference stacks. The strengths visible now are the EXL3 weight-handling path and the kernel verification design, and a performance advantage still needs separate evidence.