The fast, open-source container engine for macOS, with native Apple Silicon GPU, Neural Engine, and PyTorch acceleration.

lighter is an open-source project sponsored by Fieldwork, providing dedicated engineering time to build and maintain high-performance virtualization and AI infrastructure for Apple Silicon.

lighter is a container engine for the Mac, built from scratch in Rust on Apple's Hypervisor.framework. It is the fastest way to run Docker on Apple Silicon, and the only one that gives Linux containers Metal, the Neural Engine and the Mac's hardware video encode and decode.

A drop-in replacement for Docker Desktop, OrbStack and Colima:

• 🔄 Drop-in Docker: your existing docker, docker compose, kind and CI scripts work unchanged.
• ⚡ Fast: a cold start to a running container in 771 ms; shared folders that read faster than the Mac's own disk; installs on a shared folder as fast as on the Mac itself.
• 🚀 Apple Silicon acceleration: local LLMs on Metal, PyTorch on MPS, ONNX models on the Neural Engine and H.264/HEVC on the media engine, all from inside a container.
• 🏠 Home lab ready: Frigate supports lighter upstream, Home Assistant's voice pipeline runs on the Mac's GPU, Zigbee and Z-Wave sticks plug straight into your containers, and Pi-hole serves your whole network.
• 🧩 x86-64 images too: linux/amd64 containers run under Apple's Rosetta.
• 🪶 Light: 687 MiB with a container running (Docker Desktop: 3.4 GB), and memory goes back to macOS within a minute of work finishing.
• 🆓 Free and open source: MIT / Apache 2.0, no subscriptions, no seat licences, no telemetry, no GUI.

Requires Apple Silicon and macOS 15+ (Sequoia, Tahoe).

brew tap fieldwork-ai/tap
brew trust fieldwork-ai/tap
brew install lighter

Homebrew 7 loads formulae only from taps you have trusted, so brew trust is needed once, before the first install or upgrade from this tap.

Or with the one-line installer:

curl -fsSL https://raw.githubusercontent.com/fieldwork-ai/lighter/main/scripts/install.sh | sh

Then:

lighter start                               # boots the machine and sets up your Docker context
docker run --rm alpine echo "hello from lighter"

lighter status      # machine state, CPUs, memory and disk
lighter doctor      # check entitlements, Rosetta, accelerators and disk
lighter config      # CPUs, memory, disk, devices
lighter install     # start at login (launchd)
lighter upgrade     # move to the latest release

Direct installs can download updates in the background (lighter update auto-download on); Homebrew installs are managed by brew. See updates.

Seven ways to run a container on a MacBook Pro (M5 Pro, 18 cores, 48 GB, macOS 26), every runtime at 8 vCPUs and 16 GiB, medians of three, recorded with benchmarks/compare.sh on lighter 0.10.0. The method, the raw results and every caveat are in the record.

On a folder shared from the Mac

On the runtime's own disk

Engine

† Apple container runs a VM per container, and none once it exits.

lighter keeps a container's file cache for 30 s after it stops, so the next command reads what the last one wrote from memory; that is its higher peak, and it has given the memory back a minute later.

Media and an LLM

ᵐ on the Mac's media engine, with output identical to the Mac's own; every other runtime encodes in software.

An LLM on the GPU (llama.cpp, Qwen2.5 0.5B Q4_K_M, tokens a second)

OrbStack, Docker Desktop, Colima and Apple container give a container no GPU.

Five devices, each a standard Docker CDI device (--device lighter.sh/<name>=all), on by default. The guide has examples, limits and how each works.

docker run --rm --device lighter.sh/metal=all -v ./models:/models llama-cpp-rpc \
  llama-bench -m /models/qwen2.5-0.5b-instruct-q4_k_m.gguf --rpc "$LIGHTER_METAL" -ngl 99

• Frigate supports lighter upstream (frigate#24453, in Frigate's next release): its onnx detector runs on the Neural Engine, and preset-apple-silicon-h264 / -h265 decode cameras on the media engine. On an 8 GB M1, one detector keeps up with 14 cameras (YOLOv9-s), against 3 for Frigate's ZMQ detector on the Mac. For today's Frigate, examples/frigate-ane adds the detector to Frigate's own image.

• Home Assistant runs as it does anywhere else, and with LAN mode it discovers your devices by itself: Home Assistant on your Mac is a five-step guide. Its voice pipeline's speech-to-text runs on the Mac's GPU: examples/whisper-metal is a Wyoming whisper service that transcribes an 11 s clip in 1.3 s, against 5.8 s on the CPU.

• Host networking and discovery: a network_mode: host container is reachable from the Mac like a published port. For device discovery (mDNS and SSDP), LAN mode gives the machine its own address on your network: lighter config --lan on, then lighter restart. No sudo and no helper since 0.12.5, which holds Apple's VM Networking entitlement; a build of your own still needs sudo lighter lan enable once. See the 0.12.0 release notes.

• Zigbee and Z-Wave sticks: lighter usb attach <vendor:product> hands a stick plugged into the Mac to the guest, under the /dev/serial/by-id name Linux gives it. See the USB guide.

• Pi-hole and AdGuard Home publish DNS on port 53 (-p 53:53/udp -p 53:53/tcp); point your router at the Mac and every device on the network is covered. Published ports pass on each caller's own address, so Pi-hole shows your devices one by one (set its listening mode to all origins: FTLCONF_dns_listeningMode: all).

docker run --platform linux/amd64 … runs under Apple's Rosetta, the translator that runs Intel apps on the Mac, with x86's memory ordering switched on per thread in the guest's kernel. lighter rosetta --install fetches Rosetta if the Mac doesn't have it, and lighter doctor checks it. See x86-64.

• Docker and Compose: a registered Docker context; docker, docker compose and third-party tooling work unchanged.
• Kubernetes with kind: single- and multi-node clusters with kind, kubectl and helm. See the Kubernetes guide.
• Containers by address and name: with lighter config --direct on, the Mac reaches every container at its own IP, or as web.lighter.local (db.shop.lighter.local for a Compose service), with no port published and no root. See the 0.13.0 release notes.
• Ports and IPv6: published ports (-p 8080:80) bind on the Mac; IPv6 wherever the Mac has it.
• Shared folders that behave: changes on either side seen in milliseconds.
• Cooperative resources (experimental): lighter config --resources cooperative uses every core and grows memory as containers need it, up to twice the Mac's RAM, giving it back when they stop. See the 0.10.0 release notes.
• Headless: a background daemon or launchd service; no menu bar, no Electron.

How it is built, and why it is fast: architecture.

Needs Rust 1.85+, Docker (for the guest kernel and rootfs) and the macOS 15+ SDK.

make guest            # guest kernel and rootfs
make gpu ane metal    # accelerator libraries (optional)
make build            # the CLI and VMM, ad-hoc signed with the hypervisor entitlement
make gates            # the milestone gates: real VMs, end to end

Dual-licensed under MIT or Apache 2.0, at your option, except the guest kernel patches (guest/kernel/patches), which are under GPL-2.0, and the crates vendored in third_party/ (v4l2r, MIT; virtio-media, BSD-3-Clause).

Release archives include third-party notices in THIRD-PARTY-NOTICES, and complete source for the Linux guest is published with each release as lighter-<version>-guest-source.tar.

See CONTRIBUTING.md for contribution terms and guidelines.