Installation
Mold no longer publishes new versions to crates.io. Existing registry versions are historical; use GitHub releases, Nix, Docker, AUR, or a source build for current versions.
Native apps
- macOS (recommended): download Mold Studio for Mac, the signed, notarized native app (Apple Silicon, macOS 26 or newer), then follow the Mold Studio for Mac guide.
- macOS desktop (legacy): the Tauri Mold DMG is still published for Macs that cannot run macOS 26; see the Desktop App guide.
- Windows desktop (nightly): download the self-signed NSIS installer and its public certificate, then follow the Windows desktop trust instructions.
- iPhone: the current remote-only app is distributed through the project's invited internal and external TestFlight groups; there is not yet a public App Store listing. See the iPhone App guide for supported workflows and host setup.
- Android (nightly): download the raw signed APK directly to an Android 7.0 or newer device. No archive extraction is needed; follow the Android App guide for the one-time sideload permission.
- Linux desktop: build the Nix package for the host GPU with
nix build .#mold-desktop,.#mold-desktop-sm86, or.#mold-desktop-sm120.desktop-buildcan produce an AppImage on a conventional Linux host; tagged releases do not publish that AppImage yet. See the Desktop App guide.
Mold is CLI-native. The command-line installation below installs the primary Mold interface and the same engine/server used by the native apps, scripts, agents, and custom API clients.
Windows CLI
The current prebuilt x64 Windows CLI is a self-signed nightly CPU inference and remote-client build. Download and unpack it with PowerShell:
$release = 'https://github.com/utensils/mold/releases/download/latest'
Invoke-WebRequest "$release/mold-x86_64-pc-windows-msvc-cpu.zip" `
-OutFile mold-windows.zip
Expand-Archive .\mold-windows.zip `
-DestinationPath "$env:LOCALAPPDATA\Mold\bin" `
-Force
& "$env:LOCALAPPDATA\Mold\bin\mold.exe" versionThe zip includes mold-windows-self-signing.cert.cer. Verify its SHA-1 thumbprint is E8DA2990155CCC6E9278A8319008A763AC5DFC79 before trusting it; the Desktop App guide has the exact PowerShell trust commands. Add %LOCALAPPDATA%\Mold\bin to your user PATH for a permanent mold command. This published CLI does not include CUDA; use it for CPU work or point it at a GPU host with mold run --host http://gpu-host:7680 "prompt" or MOLD_HOST. Build from source on an x64 machine with the CUDA toolkit to get a CUDA-enabled local binary; the Windows contributor commands cover the supported helper. Upgrade by downloading and expanding the newest zip again; mold update cannot replace a running Windows executable yet.
One-Line Install (recommended)
Stable release:
curl -fsSL https://raw.githubusercontent.com/utensils/mold/main/install.sh | shDownloads the latest tagged release from github.com/utensils/mold/releases/latest and installs it to ~/.local/bin/mold. On Linux x86_64, automatic selection installs a GPU-free CPU/remote CLI when nvidia-smi is missing, fails, returns no inventory, or no GPU is visible. Otherwise it queries every GPU visible through CUDA_VISIBLE_DEVICES and selects a compatible release binary independent of device order. It supports any homogeneous device count, including RTX 3050/30-series, RTX 50-series, and named RTX PRO variants. macOS builds include Metal support.
A client needs no GPU, NVIDIA driver, or CUDA toolkit. Force that build even on a GPU machine with MOLD_BACKEND=cpu, then set MOLD_HOST=http://gpu-server:7680 (or pass --host) to generate remotely. The CPU archive does not support local GPU generation; it still supports model management, host commands and completions. Automatic selection does not check CUDA user-space libraries: on a GPU machine missing those libraries, choose CPU explicitly for remote use.
GPU-free archives become available on nightly after the first successful main publication containing this change, and on stable/AUR with the next tagged release. Older tags have no CPU archive: installation fails with guidance to use a newer published release or build from source without GPU features.
The installer downloads the exact release's SHA256SUMS and verifies the selected archive before extraction. Missing sm86, sm100, or sm120 artifacts fail closed; a higher compute target is never substituted for a lower-capability GPU. Only sm89 may use an old release's unsuffixed archive, which was the historical name for that same target. Timeouts, authentication failures, server failures, duplicate checksums, and checksum mismatches stop installation.
Nightly CLI
Install the latest successfully published CLI build from main:
curl -fsSL https://raw.githubusercontent.com/utensils/mold/main/install.sh | MOLD_CHANNEL=nightly shNightly uses the rolling latest prerelease and the same platform detection and SHA-256 verification as stable. It may be less tested than a tagged release. Re-run that command or use mold update --nightly to refresh it.
Options
All options are passed as environment variables:
# Install to a custom path
curl -fsSL ... | MOLD_INSTALL_DIR=/usr/local/bin sh
# Pin to a specific release tag (default: latest)
curl -fsSL ... | MOLD_VERSION=v0.10.0 sh
# Install the latest rolling build from main
curl -fsSL ... | MOLD_CHANNEL=nightly sh
# Force the GPU-free remote CLI (Linux x86_64)
curl -fsSL ... | MOLD_BACKEND=cpu sh
# Require CUDA rather than falling back to CPU when no GPU can be inspected
curl -fsSL ... | MOLD_BACKEND=cuda sh
# Force a GPU architecture (default: auto-detect on Linux)
curl -fsSL ... | MOLD_CUDA_ARCH=sm86 sh # Ampere (RTX 3090 / A40)
curl -fsSL ... | MOLD_CUDA_ARCH=sm89 sh # Ada (RTX 40-series)
curl -fsSL ... | MOLD_CUDA_ARCH=sm100 sh # Datacenter Blackwell (B200 / B300)
curl -fsSL ... | MOLD_CUDA_ARCH=sm120 sh # Consumer Blackwell (RTX 50-series)Note: the env var has to be on the
shside of the pipe; withVAR=value curl ... | sh, the variable only applies tocurland the installer itself still sees the default.
MOLD_BACKEND accepts auto (default), cpu, or cuda on Linux. cpu skips all GPU probing and conflicts with a nonempty MOLD_CUDA_ARCH. auto with a nonempty architecture override means explicit CUDA intent; cuda without an override requires an inspectable compatible GPU. An explicit architecture can target another machine when the inventory is unavailable or all GPUs are hidden. Malformed inventories, invalid visibility selectors and unsupported/mixed fleets still fail; select cpu explicitly for a remote client. Non-auto backend values are rejected on macOS, whose archive includes Metal.
An explicit MOLD_CUDA_ARCH must equal the target selected for every visible GPU. Homogeneous 8.6 and 8.9 fleets use sm86 and sm89 respectively. A mixed 8.6/8.9 fleet uses sm86: release CI requires exact sm86 PTX embedded in the final executable, which both Ampere and the forward-compatible Ada driver can JIT. Homogeneous 10.x and 12.x fleets use sm100 and sm120. Compute capability 8.0, 9.x, and unproven mixed families fail closed because no release tarball is qualified for their floor. Narrow CUDA_VISIBLE_DEVICES or build from source with verified targets.
PTX compatibility is forward-only toward equal-or-higher device compute capabilities. An sm86 artifact can JIT on sm86 or sm89, but sm89 PTX cannot JIT backward on an RTX 3090 (sm86). The RTX 3090 distribution gate therefore requires a successful sm86 generation and treats the same-source sm89 failure as an expected incompatibility regression, not a successful smoke.
MOLD_VERSION accepts any tag that exists on the releases page, for example v0.8.0 to reproduce an older install. Without it the script follows the releases/latest redirect on GitHub and installs whatever that currently points at.
Updating
mold update preserves the compiled backend: the GPU-free Linux CLI downloads only the CPU archive without probing NVIDIA; CUDA builds retain architecture selection. To change backend, rerun the installer with MOLD_BACKEND set. A missing CPU archive never falls back to CUDA or a different release.
mold update # Update to latest release
mold update --nightly # Install latest rolling build from main
mold update --check # Check for updates without installing
mold update --version v0.7.0 # Install a specific versionOr re-run the install script:
curl -fsSL https://raw.githubusercontent.com/utensils/mold/main/install.sh | shArch Linux / AUR
Three packages on the AUR:
paru -S mold-ai-bin # Prebuilt GPU-free CLI for remote hosts; no CUDA required.
paru -S mold-ai # Builds from source; set CUDA_COMPUTE_CAP for other GPUs
paru -S mold-ai-git # Builds from main HEADSubstitute yay, pikaur, or any other AUR helper as appropriate. With vanilla makepkg:
git clone https://aur.archlinux.org/mold-ai-bin.git
cd mold-ai-bin
makepkg -siConflict with extra/mold: All three packages declare conflicts=('mold') because they install /usr/bin/mold; the same path used by the rui314 linker. You cannot have both installed simultaneously. If you need the linker for your build toolchain, install mold via Nix or the one-line installer (which targets ~/.local/bin) instead.
Migration: mold-ai-bin now installs the GPU-free CPU archive instead of SM89 CUDA. Existing users who generate locally must switch to mold-ai or mold-ai-git, or install a matching CUDA release archive. Source packages build against Arch's installed CUDA runtime and still require an NVIDIA GPU for local generation. Use an explicit compute capability:
CUDA_COMPUTE_CAP=89 paru -S mold-ai # RTX 40-series
CUDA_COMPUTE_CAP=86 paru -S mold-ai # RTX 3090 / A40
CUDA_COMPUTE_CAP=100 paru -S mold-ai # B200 / B300
CUDA_COMPUTE_CAP=120 paru -S mold-ai # RTX 50-seriesThere is no mold-ai-bin-sm100 package before real B200 qualification.
The GPU-free binary package has no CUDA runtime dependency, fixing the libcudart.so.12 loader failure on client machines. CUDA release archives still need their matching CUDA 12 user-space libraries; installing Arch's CUDA 13 toolkit alone does not satisfy those archives.
To upgrade: paru -Syu mold-ai-bin (or mold-ai / mold-ai-git). mold update will detect a pacman-managed install and direct you here instead of attempting to overwrite the binary.
To uninstall: sudo pacman -R mold-ai-bin (or whichever package you installed).
Nix
The flake is a source revision channel, not the self-updater's binary release channel. Pin the flake input or Git revision for reproducibility. Nix packages do not consult MOLD_DISTRIBUTION_IMAGE_VERSION; source and Nix-built cloud clients default to the mutable latest* container channel unless a release builder explicitly embeds an official distribution version.
# Run directly: no install needed
nix run github:utensils/mold -- run "a cat"
# RTX 3090 / A40
nix run github:utensils/mold#mold-sm86 -- run "a cat"
# B200 / B300
nix run github:utensils/mold#mold-sm100 -- run "a cat"
# RTX 50-series
nix run github:utensils/mold#mold-sm120 -- run "a cat"
# Add to your system profile
nix profile install github:utensils/moldFrom Source
Source builds likewise do not imply that a same-version GHCR image exists. They use rolling latest* cloud images by default. For local CUDA compilation, set CUDA_COMPUTE_CAP to the target required by the GPUs you intend to expose.
./scripts/ensure-web-dist.sh && cargo build --profile dev-fast -p mold-ai --features cuda,flash-attn./scripts/ensure-web-dist.sh && cargo build --profile dev-fast -p mold-ai --features metalcargo build --release -p mold-ai --features cuda,flash-attn,cudnncargo build --release -p mold-ai --features metalRequires Rust 1.93+ and CUDA toolkit (Linux) or Xcode (macOS).
Name flash-attn on a CUDA build. FLUX.1 and FLUX.2 changed their arithmetic in mold 0.29, and the change lands whether or not the kernel is compiled: their math path folds the softmax scale into K either way. A CUDA binary without flash-attn is therefore correct, and slower than one with it, for exactly the same pixels — the worst of both. Every shipped Linux CUDA artifact compiles it except mold-sm120, which stays on math attention until FlashAttention's tile selection is measured on an RTX 50-series card, and on that hardware you should leave flash-attn off too. cudnn is the matching convolution backend and needs the cuDNN SDK's headers, which is why it is not implied by cuda; drop it if you do not have them. Adding the kernels costs several minutes of nvcc on the first build.
Optional features can be added to the same build, for example --features cuda,flash-attn,preview,expand,discord or --features metal,preview,expand,discord if you also want terminal preview, local prompt expansion, or the Discord bot.
mesh-texture builds the pinned xatlas CPU UV unwrapper. Nix CLI and desktop recipes include it; source builds need a C++11 compiler. It does not by itself enable a texture engine or bypass the recipe’s advertised capabilities.
pulid (face identity) is one of those features, and it is the only one with a build-time dependency of its own: protoc must be on PATH, because candle-onnx's build script drives prost-build. It is in the nix develop shell already; otherwise install it before building, since the build fails partway through without it.
brew install protobuf # macOS
sudo apt-get install -y protobuf-compiler # Debian/Ubuntu
sudo pacman -S protobuf # Arch
sudo dnf install protobuf-compiler # FedoraEvery official binary (the release tarballs, the Nix packages, the AUR packages, and the desktop app) already ships with pulid on, so this only concerns building it yourself.
dev-fast is the repo's local-iteration profile: it keeps debuginfo, enables incremental compilation, and uses thin LTO plus more codegen units so optimized builds stay much faster than the shipping --release profile.
Docker
# Ada / SM89 only; choose an architecture-specific tag for other GPUs
docker pull ghcr.io/utensils/mold:latest
docker run --gpus all -p 7680:7680 ghcr.io/utensils/mold:latestSee Docker & RunPod for full deployment instructions.
Pre-Built Binaries
The one-line installer targets the latest stable tag by default; with MOLD_CHANNEL=nightly, it targets the rolling latest prerelease. Both use assets from the releases page:
| Platform | File |
|---|---|
| macOS Apple Silicon | mold-aarch64-apple-darwin.tar.gz |
| Linux x86_64 (Ampere, RTX 3090 / A40) | mold-x86_64-unknown-linux-gnu-cuda-sm86.tar.gz |
| Linux x86_64 (GPU-free remote CLI) | mold-x86_64-unknown-linux-gnu-cpu.tar.gz |
| Linux x86_64 (Ada, RTX 4090 / 40-series) | mold-x86_64-unknown-linux-gnu-cuda-sm89.tar.gz |
| Linux x86_64 (datacenter Blackwell, B200 / B300) | mold-x86_64-unknown-linux-gnu-cuda-sm100.tar.gz |
| Linux x86_64 (consumer Blackwell, RTX 50-series) | mold-x86_64-unknown-linux-gnu-cuda-sm120.tar.gz |
B200 support is simulated, not hardware-qualified. The sm_100 artifact passes hosted build, CUDA-image, loader, NVML, archive, and synthetic scheduler checks; real 8×B200 and MIG qualification remain deferred. GH200, GB200, and GB300 require future linux/arm64 artifacts and are unsupported. Current Linux release archives and containers are amd64-only.
To install an older tag, put MOLD_VERSION=<tag> on the sh side of the pipe, e.g. curl -fsSL ... | MOLD_VERSION=v0.8.0 sh. Placing it on the curl side (VAR=value curl ... | sh) exports the variable to curl only; the installer still sees the default and installs the latest release.
Pinned releases remain pinned. If an older tag lacks the selected native sm86, sm100, or sm120 asset, the installer and mold update --version fail closed. An old unsuffixed CUDA archive is considered only when the selected target is sm89, because that was the former filename for the same target.
Shell Completions
source <(mold completions bash) # bash
source <(mold completions zsh) # zsh
mold completions fish | source # fishThe Linux CUDA release archives also ship these scripts pre-generated beside the binary, as completions/mold.bash, completions/_mold and completions/mold.fish, so a packager never has to execute the CUDA-linked binary to obtain them. The AUR packages install them system-wide.