Hawking Runtime is a local inference and GPU computing project for Apple Silicon Macs. It has two components, Extension and eGPU. They are parts of one project and can be used independently.

Extension

Runs local models on Apple Silicon Macs and makes them available to applications and coding agents through an OpenAI-compatible API on 127.0.0.1:8277. Requires macOS on Apple Silicon and Python 3.12.

It checks the available compute paths, runs models with a verified, pinned llama.cpp build (the default engine; an experimental MLX adapter is available for supported configurations) and serves them with Chat Completions, streaming, tool calls and /v1/responses. Existing engines on other hosts can be registered over authenticated HTTPS and used according to their measured capacity. The Python package does not contain the NVIDIA driver or CUDA backend.

Installation, usage, the coding-agent setup guide (AGENT_INSTALL.md) and troubleshooting: extension/ in the Hawking Runtime repository ↗.

eGPU

Developer source for running an NVIDIA GPU over Thunderbolt on Apple Silicon macOS: a DriverKit driver, host runtime, CUDA kernel catalogue, diagnostics and build tooling, with a public C ABI. The reference setup is an NVIDIA RTX 4000 SFF Ada on an M3 Ultra Mac Studio.

Demonstrated means it ran on the maintainers' qualification build. Public status is what the release is qualified for.

CapabilityDemonstrated (qualification build)Public release status
Driver initializationYesDeveloper source; not qualified for public use
Kernel execution and numerical verificationYes. SGEMM about 10 GFLOP/s through host-coherent memory.Developer source; not qualified for public use
Recovery without rebootYesNot qualified
PCIe link speedGen2 x4 measured at about 1.9x Gen1 transfer in a single measured trial; Gen3 not achieved on the reference enclosure.Gen1 qualified; Gen2 opt-in, not qualified.
Public C ABI computeYes, on the reference setup at Gen1: 61 of 61 known-answer checks (vecadd, grid-stride vecadd, SGEMM exact and random).Qualified on the reference setup (Gen1).

Build, signing and lifecycle instructions: egpu/ in the Hawking Runtime repository ↗ (README, SIGNING.md, docs/LIFECYCLE.md). The Extension Python package does not contain the NVIDIA driver or CUDA backend.

Compatibility and limits

Supported local configuration: Apple Silicon Mac, Python 3.12. The local server binds to loopback only. Hawking refuses a model whose memory plan does not fit; free memory, choose a smaller model, or run hawking serve MODEL --dry-run first. The eGPU component is developer source and requires hardware you configure and build yourself.

Source: github.com/hawking-systems/runtime ↗ (extension/, egpu/). Package: hawking-runtime on PyPI ↗. Published with the initial release.