mold

Local AI image and video generation on your own GPU — NVIDIA CUDA and Apple Silicon Metal, no Python, no cloud account, no usage fees. CLI-native and pipe-friendly, with a native desktop app, web studio, TUI, iPhone and Android companions, Discord bot, and REST/SSE API built on the same engine.

Documentation · Models · Desktop guide · API

Install

Stable release:

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

Nightly CLI from the latest published main build:

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

The installer picks the right prebuilt binary for your GPU and verifies its checksum. Use mold update to stay on stable or mold update --nightly to install the newest nightly. Nix (nix run github:utensils/mold), Arch (paru -S mold-ai-bin), and source builds are covered in the installation guide; binaries and checksums are on the releases page. GH200, GB200, and GB300 require future linux/arm64 artifacts and are unsupported.

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

• Models: FLUX.1, Flux.2 Klein/Dev, SD 1.5, SDXL, SD 3.5, Z-Image, Qwen-Image, Qwen-Image-Edit, Wuerstchen v2, LTX Video, LTX-2 / LTX-2.3, Wan 2.1/2.2, and MiniMax H3 — see the model catalog for sizes, VRAM needs, and settings
• Images: text-to-image, img2img, multimodal editing, inpainting, ControlNet, LoRA, prompt expansion, and Real-ESRGAN upscaling
• Face identity (PuLID): keep one person's face across arbitrary prompts with --id-image (repeatable up to 4, averaged into one identity), on FLUX (flux-dev:q4, flux-dev:q8, the pulid-flux bundle) and SDXL (sdxl-base:fp16, juggernaut-xl:fp16, realvis-xl:fp16, dreamshaper-xl:fp16, the pulid-sdxl bundle) — pure Rust SCRFD, ArcFace, a BiSeNet face mask, EVA02-CLIP, and IDFormer shared by both adapters, feeding twenty cross-attention modules inside the FLUX transformer or seventy inside the SDXL UNet, plus --true-cfg on FLUX for a real negative branch on an otherwise guidance-distilled model (SDXL's own --guidance is already the classifier-free scale)
• Video and audio: text/image-to-video, multi-prompt sequences, clip continuation (--extend), lip dub (--pipeline lip-dub), text-to-audio (--pipeline t2a), native MP4 with generated audio, and LTX-2 output up to 4K via tiled composition
• Fits your hardware: quantized variants, encoder fallback, smart VRAM placement, block offloading, and spatial tiling (--spatial-tile)
• Multi-machine: connect LAN/Tailscale hosts and RunPod, route jobs by capability, and browse every machine's gallery in one place
• Library organization: title (--title), favorite, tag, and collect prints — or file them at creation with --tag / --collection so they arrive organized — with a per-host trash and configurable retention (gallery.trash_retention_days, mold trash) instead of permanent delete — merged across machines in the web and desktop Library (Prints | Collections | Trash)

MiniMax H3 weights use the MiniMax H3 Community License, not Mold's MIT license. H3 may be used through Mold in every territory and workflow — local, remote, shared, hosted, output distribution, and redistribution — with no separate acceptance step; review the linked terms for your use. The reviewed FL2VA Turbo distillations are ordinary model tags (minimax-h3-fl2va:comfy-pruned-int8-turbo-8step and …-turbo-4step-768p) that pull the same compact stack plus one pinned LoRA adapter and render at their tier's fixed step count. The compact NVFP4 transformer tags (minimax-h3-fl2va:comfy-pruned-nvfp4 and minimax-h3-ref2va:comfy-pruned-nvfp4) download and verify but have no runtime yet. Current capability limits (FL2VA on SM89 CUDA only) are documented in the H3 model guide.

Mold Studio

One native desktop app for macOS, Linux, and Windows with five workspaces — Create, Library, Models, Machines, and Settings — spanning local and remote generation, a merged multi-machine gallery, model discovery from Hugging Face and Civitai, GPU telemetry, and QR pairing for the iPhone companion.

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

Android uses the same remote-only Mold Studio mobile interface. Download the signed universal APK directly—there is no zip to unpack:

Download stable Android APK · Download nightly Android APK · Android installation guide

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 terminal interface with mold tui:

The TUI Create workspace with a native terminal image preview

Run the engine where the GPU lives and point any client at it:

mold serve                                      # GPU machine
MOLD_HOST=http://gpu-server:7680 mold run "a cat"  # laptop

See the remote workflow and RunPod guides.

Install Mold's embedded Agent Skill for your coding agent:

mold skill install --detected

mold skill list shows all supported agents and paths; explicit targets such as mold skill install claude codex and project installs with --project are also supported.

Project

Mold is a Rust workspace built on candle. The documentation covers the CLI, configuration, deployment, and HTTP API.

Core contributors: James Brink and Jeffrey Dilley.

Licensed under the MIT License.

Third-party code. The LTX-Video transformer, 3D causal VAE, and flow-match scheduler (crates/mold-candle/src/ltx_video/), and the LTX-2 video transformer and VAE derived from them, were ported from candle-video by FerrisMind (Copyright 2025 FerrisMind), licensed under the Apache License 2.0 — itself a Rust port of Hugging Face diffusers. Those files remain Apache-2.0; see THIRD_PARTY_NOTICES.md for this and every other third-party notice.

Face-identity weights. Face identity additionally downloads two InsightFace pretrained models (scrfd_10g_bnkps, glintr100), which are licensed for non-commercial research purposes only — the InsightFace code is MIT, the weights are not. Mold ships neither and refuses to download them until you record acceptance with mold pull pulid-flux --accept-license insightface-antelopev2 (or pulid-sdxl — the acceptance covers both bundles); mold licenses lists what has been accepted. The PuLID adapters are Apache-2.0, the EVA02-CLIP tower is MIT, and facexlib's BiSeNet face parser (parsing_bisenet.pth, masking the aligned crop before the tower sees it) is also MIT with no acceptance step of its own — each bundle is five artifacts in total: about 2.3 GB for pulid-flux, about 2.2 GB for pulid-sdxl.