mold

Local AI image and video generation on your own GPU. Mold runs on NVIDIA CUDA and Apple Silicon Metal with no Python, no cloud account, and no usage fees.

Mold began at the command line and remains CLI-native: every core workflow is available as a composable command with predictable stdin, stdout, files, and exit behavior. Use it interactively, script it in a pipeline, give it to an agent, run it in CI, or build on the same engine through REST and SSE. The desktop, web, TUI, and iPhone experiences are additional interfaces—not replacements for the CLI foundation.

Documentation · Models · Desktop guide · API

Mold Studio

Mold Studio brings local and remote generation into one native desktop app for macOS and Linux:

• Create images, edits, upscales, and multi-stage videos with capability-aware controls, prompt expansion, batches, LoRAs, ControlNet, and reproducible seeds.
• Library merges prints from every connected machine into a fast, searchable gallery with saved prompts, settings, provenance, and native media actions, multi-select on every GUI surface, delete-across-devices semantics, and a remembered thumbnail-size control on web and desktop.
• Models discovers and installs checkpoints from Mold, Hugging Face, and Civitai, with live download progress and per-machine routing.
• Machines connects this device, LAN/Tailscale hosts, and RunPod while showing GPU telemetry, queues, downloads, and installed models.
• Settings includes Mold and Safelight themes, Stable/Nightly updates, local performance controls, and secure provider credentials.

Download Mold for macOS (Apple Silicon) · Explore the desktop app

Mold also ships a responsive web studio with every mold serve, a keyboard-first terminal UI, a REST/SSE API, and a remote iPhone companion. On CUDA servers, resource telemetry is joined to CUDA's startup-visible devices by UUID. CUDA_VISIBLE_DEVICES remains a hard boundary even when it reorders cards or uses GPU/MIG UUID selectors: hidden GPUs are not published, and a MIG worker never inherits its parent GPU's full-memory sample. One versioned device registry supplies those telemetry targets, worker construction, scheduler candidates, /api/devices, and legacy status, so a device cannot silently acquire a different identity between subsystems. GPU startup defaults to all. --gpus / MOLD_GPUS also accepts none (maintenance mode), visible ordinals, Mold stable IDs (cuda:<uuid> or metal:default), and NVIDIA GPU-... / MIG-... UUIDs. Prefer stable IDs in persistent configuration because ordinals are process-local. Maintenance mode keeps inventory, telemetry, downloads, and settings available but rejects generation and model-load requests. Use mold gpu list, mold gpu disable <ID>, and mold gpu enable <ID> for runtime administration when /api/capabilities.devices.lifecycle is true. That flag is true only while authoritative scheduler V2 owns live GPU workers; legacy, observe, CPU-fallback, and maintenance runtimes are read-only and reject lifecycle mutation. Busy devices drain their current generation before disabling; startup-excluded devices require a restart. Scheduler admission uses each device's sampled free VRAM rather than total capacity and freezes a concrete per-component execution plan before dispatch. Explicit CPU/device placement is a hard constraint; missing devices and cross-GPU component pins fail instead of falling back. mold run --local --batch N uses the same deterministic assignment core across every GPU selected by --gpus/MOLD_GPUS; single-item local runs retain best-free-GPU selection.

Web, desktop, and iPhone preview each finalized ordinary-generation request on candidate hosts before queueing. The read-only preview preserves the sibling count without reserving a GPU or loading weights, redacts prompt/media content, and keeps prepared work pinned to the same endpoint, key, and server instance. A strict non-authoritative unsupported response or legacy 404/405 keeps backward-compatible routing; infeasible, malformed, authentication, upgrade, and server failures queue nothing. Exact chain and local prompt-expansion/ post-generation-upscale utility previewing is deliberately deferred rather than guessed.

Scheduler V2 learns cold load, warm reload, and execution time separately, so ETA planning adds the applicable setup cost exactly once. Machine detail, status, TUI, CLI, MCP, and Discord surfaces report every device; interactive device buttons appear only for authoritative live lifecycle control, with the supported restart-enable recovery shown for read-only dispatch modes.

Scheduler V2 is the default GPU dispatch owner. During the one-release rollback window, restart with MOLD_DISPATCH_MODE=legacy to restore the previous depth-two dispatcher or MOLD_DISPATCH_MODE=observe to keep legacy dispatch authoritative while recording read-only V2 decisions. The mode is restart-only; Mold never switches CUDA owners or contexts live. Queue pause applies to generation, utility, and admin GPU work in every mode.

Install

curl -fsSL https://raw.githubusercontent.com/utensils/mold/main/install.sh | sh

The installer downloads the latest release to ~/.local/bin/mold, verifies its SHA-256 checksum, inspects every Linux GPU visible through CUDA_VISIBLE_DEVICES, and selects one compatible NVIDIA build independent of device order. Unsupported mixed architecture groups fail with an actionable override/source-build error instead of silently targeting GPU 0. macOS installs the Metal build.

Other install options

# Nix — NVIDIA Ada / RTX 40-series
nix run github:utensils/mold -- run "a cat"

# Nix — NVIDIA Ampere / RTX 3090 or A40
nix run github:utensils/mold#mold-sm86 -- run "a cat"

# Nix — NVIDIA datacenter Blackwell / B200 or B300
nix run github:utensils/mold#mold-sm100 -- run "a cat"

# Nix — NVIDIA consumer Blackwell / RTX 50-series
nix run github:utensils/mold#mold-sm120 -- run "a cat"

# Arch Linux
paru -S mold-ai-bin

# Build from source on Linux
./scripts/ensure-web-dist.sh
cargo build --release -p mold-ai --features cuda

# Build from source on Apple Silicon
./scripts/ensure-web-dist.sh
cargo build --release -p mold-ai --features metal

B200/B300 sm_100 builds and synthetic multi-device coverage are simulated, not hardware-qualified. Real 8×B200 qualification remains deferred. GH200, GB200, and GB300 require future linux/arm64 artifacts and are unsupported. The deferred RTX 3090 acceptance runner and evidence schema live in docs/qualification/; no passing report is claimed.

Prebuilt binaries and checksums are available on the releases page. See the installation guide for platform details.

Quick start

# Generate with the default model
mold run "a cat riding a motorcycle through neon-lit streets"

# Choose a model and reproducible seed
mold run flux-dev:q4 "a sunset over mountains" --seed 42

# Edit an image
mold run qwen-image-edit-2511:q4 "make the chair red" --image chair.png

# Generate video
mold run ltx-video-0.9.6-distilled:bf16 "a fox in the snow" --frames 25

# Launch the web studio and API
mold serve

Models download automatically on first use. Generated media is saved locally with prompt, model, seed, and generation metadata.

What it supports

• FLUX.1, Flux.2 Klein and Dev, SD 1.5, SDXL, SD 3.5, Z-Image, Qwen-Image, Qwen-Image-Edit, Wuerstchen v2, LTX Video, and LTX-2 / LTX-2.3
• Text-to-image, image-to-image, multimodal editing, inpainting, ControlNet, LoRA, prompt expansion, and Real-ESRGAN upscaling
• Text/image-to-video, multi-prompt video chains, one-request continuation of an existing clip (--extend), native MP4 output, checkpoint-dependent generated audio, and LTX-2 guidance overrides across CLI, web, desktop, and iPhone
• Sequence authoring inside Create with per-clip prompts, duration, source images, Smooth / Cut / Fade seams, resumable jobs, in-place editing that re-renders only the changed clips, explicit remote-machine routing, and a web dry run that shows the server-normalized plan before anything is queued
• Quantized model variants, encoder fallback, smart VRAM placement, and FLUX block offloading
• Local CLI, native desktop, browser, TUI, iPhone, Discord, and authenticated REST/SSE clients

Browse the model catalog for sizes, VRAM requirements, and recommended settings.

More ways to create

Preview generations directly in supported terminals:

mold run "a cat" --preview

Inline image generation in Ghostty with --preview

Or open the keyboard-first Mold Studio terminal interface:

mold tui

The TUI Create workspace with a native terminal image preview

Remote and cloud GPUs

Run the engine where the GPU lives and use any Mold client over HTTP:

# GPU machine
mold serve

# Laptop or another client
MOLD_HOST=http://gpu-server:7680 mold run "a cat"

For managed cloud jobs, mold runpod can provision a GPU, reuse network volumes, stream progress, and save the result locally. See the remote workflow and RunPod guides.

Project

Mold is a Rust workspace built on candle. Its CLI-native architecture keeps the engine useful from a terminal, shell pipeline, agent, CI job, native app, or custom client. The documentation covers the CLI, configuration, deployment, and HTTP API.

Core contributors: James Brink and Jeffrey Dilley.

Licensed under the MIT License.