lmm

Local AI model manager for macOS and Linux.

Download once. Use everywhere. Remove cleanly.

Why lmm?

Local AI tools (LM Studio, Ollama, llama.cpp, ComfyUI, …) each maintain their own model directories. The same 8 GB model downloaded through different tools can consume 16–24 GB of disk.

lmm uses HF Cache as the single canonical store and exposes models to each tool via symlinks — zero duplication, instant availability.

Features

• Interactive search — browse and install Hugging Face models from the terminal
• Format support — MLX, GGUF, safetensors
• Tool adapters — LM Studio, llama.cpp, Jan, ComfyUI, text-generation-webui, and more
• Adopt existing models — track models already in your HF Cache without re-downloading
• Doctor & GC — validate integrity, consolidate external models, reclaim disk space
• Shell completions — bash, zsh, fish, elvish, powershell

Install

macOS:

brew install holotherapper/tap/lmm

From source (macOS / Linux, requires Rust 1.87+):

cargo install --git https://github.com/holotherapper/lmm

Quick Start

lmm search                                          # interactive search → select → install
lmm add mlx-community/Qwen3-8B-4bit                 # install from Hugging Face
lmm add mlx-community/Qwen3-8B-4bit --tool lmstudio # install + expose to LM Studio
lmm list                                             # see all local models (with repo column)
lmm adopt                                           # track existing HF Cache models
lmm doctor                                          # validate + consolidate external models
lmm remove qwen3-8b-4bit-mlx                        # remove and reclaim disk space
lmm remove mlx-community/Qwen3-8B-4bit              # remove by repo name
lmm remove gemma-3-1b-it-qat-4bit                   # remove external models directly

Commands

Command	Description
lmm add <repo>	Install a model and expose to tools. Aliases: a, i, install
lmm remove [names]	Remove tracked, external, or untracked models. Alias: rm
lmm list	List all local models with repo source. Alias: ls
lmm info <name>	Show details for matching models (accepts org/repo)
lmm search [query]	Search Hugging Face. Aliases: find, discover
lmm adopt [names]	Track unmanaged HF Cache models
lmm update [names]	Update models to latest revision. Alias: upgrade
lmm doctor	Validate cache, exposures, and external models
lmm gc	Clean temp files, stale entries, orphan blobs (--dry-run for preview)
lmm config get/set/path	Read or update configuration
lmm completions <shell>	Generate shell completions

Run lmm <command> --help for full options.

Key options for add

Option	Description
--file <PATTERN>	Wildcard for GGUF selection (e.g. *Q4_K_M*)
--tool <T1,T2>	Target tools (comma-separated)
--format <F>	auto, mlx, gguf, or safetensors
--all	Expose to all compatible tools
--dry-run	Preview without changes
--replace	Replace existing exposure
-y, --yes	Skip confirmation

Deletion rules for remove

Names can be aliases (qwen3-8b-4bit-mlx) or repo identifiers (mlx-community/Qwen3-8B-4bit).

Target	Behavior
Tracked (managed)	Deletes exposure + HF Cache blobs
Tracked (adopted)	Deletes exposure only (add --purge-cache for HF Cache blobs)
External (lmstudio, jan, …)	Deletes files from the tool directory
External (ollama)	Delegates to ollama rm
Untracked HF Cache	Deletes the repo cache directory (specify by org/repo)

Tool Adapters

Symlink / reference adapters (--tool)

Tool	ID	Formats	Status
LM Studio	lmstudio	MLX, GGUF, safetensors	stable
Jan	jan	GGUF	experimental
llama.cpp	llama-cpp	GGUF	stable
ComfyUI	comfyui	safetensors, MLX, GGUF	experimental
AUTOMATIC1111 / Forge	a1111	safetensors, GGUF	experimental
InvokeAI	invokeai	safetensors, GGUF	experimental
Fooocus	fooocus	safetensors, GGUF	experimental
text-generation-webui	text-gen-webui	MLX, GGUF, safetensors	experimental
GPT4All	gpt4all	GGUF	experimental

Experimental adapters appear in the TUI when their path is configured (lmm config set paths.comfyui /path/to/ComfyUI/models).

HF Cache direct (no action needed)

These tools read HF Cache directly — lmm add makes models instantly available:

mlx-lm, transformers, mflux, Diffusers, mlx-whisper, faster-whisper, Kokoro, Bark, F5-TTS, MeloTTS, sentence-transformers, AudioCraft, vLLM, KoboldCpp, Tortoise TTS

Display only

Tool	Reason
Ollama	Copies to blob store, doubling disk usage. Shown in lmm list; removable via lmm remove (delegates to ollama rm)
Draw Things	App Sandbox + proprietary format. Not manageable

Configuration

Config file: ~/Library/Application Support/dev.local.lmm/config.json (macOS) or ~/.local/share/lmm/config.json (Linux). Override with LMM_STATE_DIR.

lmm config set paths.comfyui /path/to/ComfyUI/models
lmm config set defaults.format gguf
lmm config get paths.lmstudio
lmm config path

Run lmm config set to see available keys.

Environment Variables

Variable	Description
LMM_STATE_DIR	Override state directory
HUGGINGFACE_HUB_CACHE	Override HF Cache directory
HF_TOKEN	Hugging Face API token (for gated/private models)
NO_COLOR	Disable color output

Safety

• Adopted models keep HF Cache files unless --purge-cache is explicitly used
• Exposure removal validates paths are under the tool's configured root before deletion
• --dry-run previews any destructive operation before execution
• Symlink exposures are created atomically (stage → rename)

License

Dual-licensed under either of:

• MIT license
• Apache License 2.0

at your option. SPDX identifier: MIT OR Apache-2.0.