The host GPU, inside the sandbox

The guest gets the host GPU through Vulkan on Linux, macOS and Windows, so an agent can run ML models inside its own VM.

Agents that segment video, speak text or answer questions with a local model can do that work inside the VM, on whichever host they run.

The guest GPU is Vulkan. CUDA is not available inside the sandbox, so PyTorch and vLLM cannot use the GPU there.