Cua researchers have released an experimental compatibility layer that allowed `llama.cpp` to use faster Metal kernels inside a macOS virtual machine on Apple Silicon. In their M1 Ultra tests, the change delivered sevenfold to sixteenfold gains over the same workloads in an unmodified guest and brought some results close to bare-metal speed.
Apple's Virtualization.framework presents the guest with a virtual graphics device backed by the host GPU. The guest sends Metal work through a virtualization-aware driver while the host retains hardware control. This is paravirtualization, not physical PCI or VFIO passthrough.
In the stock Tahoe virtual machine, the device reported a conservative feature profile: roughly an Apple 5-era GPU family, 32 KB maximum threadgroup memory and no SIMD-group matrix support. Applications query those capabilities to choose code paths, so `llama.cpp` selected older kernels even though the researchers found that the virtual device could execute newer ones.
Their shim runs inside one guest process and intercepts selected Metal queries, changing the reported Apple-family and threadgroup-memory values. It leaves the workload on Apple's existing virtual GPU path and does not alter the kernel or assign the physical GPU. Missing or malformed configuration returns the process to stock behavior. Each additional Metal function still requires separate validation.
TinyLlama 1.1B prompt processing improved 11.08 times and token generation 16.36 times versus the stock guest. Prompt speed reached 98 percent of bare metal. For Google's 6.98GB Gemma 4 12B QAT Q4_0 model, prompt processing rose 7.20 times and generation 14.54 times, reaching 99.59 percent and 94.82 percent of bare-metal performance respectively.
A text-only Muse Glimmer 30B test in a 64 GiB guest improved prompt processing 7.55 times and generation 8.87 times. It used `llama.cpp` without Ollama, multimodal projection or speculative drafting. The project includes source, build scripts, a capability probe and raw logs for reproduction.
The principal benchmark environment was one M1 Ultra with a 48-core GPU and macOS 26.6.1, using a Tahoe 26.5.2 guest with eight virtual CPUs and 16 GiB of memory. Results were medians from ten samples for each row.
The release demonstrates that reported capabilities, rather than raw hardware access alone, can constrain virtualized inference. It does not establish equivalent gains across other chips, guests or Metal programs. Cua labels the work a research release and invites broader testing to determine where changing the advertised profile remains both correct and beneficial.
The authors deliberately removed broader experimental hooks, including private feature profiles and ray-tracing overrides, leaving a smaller mechanism tied to the capabilities needed by their tested inference path. That narrower scope makes the claim reproducible while limiting what the release says about general virtualized graphics.
