OrcaRouter has published a community build of Qwen3.8-27B in MLX format for Apple Silicon. The package adapts the 27-billion-parameter vision-language model and reduces its refusal behavior; it is not an official new model release from Qwen.
The repository provides 2-bit, 4-bit, 6-bit and 8-bit weight variants. The 4-bit version is also stored at the repository root for applications that load a model without choosing a subfolder. OrcaRouter keeps the vision tower and several other components at higher precision while quantizing the language-model weights.
The maintainer labels the 2-bit files as an archive rather than a usable default, reporting repetition and garbled output at that compression level. Its published throughput figure was measured on an Nvidia H200 using an MLX CUDA backend, not on a Mac, so it does not establish Apple Silicon performance.
OrcaRouter lists an Apache 2.0 license and describes controlled research as the intended use, warning that the modification substantially reduces safety protections. A low refusal rate is a property of this modification, not evidence of more accurate answers or a successful safety evaluation.