omm — Open source Model Manager

omm is an apt/brew-style package manager for local GGUF models. It keeps one model hub, exposes models to supported local AI runners, checks whether a model fits the current hardware, and can verify real local generation before you rely on it.

• Installs GGUF models into a central, configurable model hub.
• Links or imports models across Ollama, LM Studio, Jan, AnythingLLM, Msty, text-generation-webui, and KoboldCpp without silently duplicating large files.
• Ranks models against live RAM, VRAM, operating-system, and runner state.
• Verifies load and generation through local Ollama or LM Studio APIs.
• Keeps telemetry, benchmark uploads, and error reports opt-in.

• What omm does
• Install
  • Windows
  • macOS
  • Linux
  • Any OS via PyPI or pipx
  • npm
  • Troubleshooting
  • Windows portable and Winget status
  • Supported platforms
  • Local AI runners
• Usage
  • Setup & discovery
  • Install & manage models
  • Verify & benchmark
  • Update & configuration
  • Scripting
• Self-hosted benchmark data
• Signed recommendation data
• Development
• Contributing
• Security
• License

Pick your OS and follow one path from top to bottom:

• Windows — PowerShell one-liner (verified Git-source installer)
• macOS — Terminal one-liner, or the Homebrew Tap
• Linux — shell one-liner
• Any OS via PyPI or pipx — pip / pipx, no signature verification
• npm — standalone native command distributed through npm
• Troubleshooting — what you see, why, and how to fix it

1. Open this app. Open PowerShell — either Windows PowerShell 5.1 (Start menu → "Windows PowerShell") or PowerShell 7 (pwsh). Windows Terminal is fine as long as the active tab is a PowerShell tab. Do not use Command Prompt (cmd.exe), and do not paste the one-liner into Git Bash or WSL; those are Unix shells and need the Linux path instead.

2. Requirements. Python 3.10+. Windows 10 22H2/11 is the supported baseline. The installer bootstraps Python and git via winget if they are missing (winget is built into Windows 10 2004+ and Windows 11 — on older Windows, install Python 3.10+ and git manually first), then installs omm through that exact validated Python interpreter. The optional NVIDIA detector is installed only when nvidia-smi indicates an NVIDIA driver.

3. Run the installer. Both lines are one command; paste them together.

# This must run before irm: script-internal TLS settings are too late for its first download.
[Net.ServicePointManager]::SecurityProtocol = [Net.SecurityProtocolType]::Tls12; irm https://omm.run/install.ps1 | iex

4. After install. Open a new PowerShell window so your PATH picks up omm, then run:

omm        # first run: setup wizard (hardware scan + engine checklist)
omm scan   # hardware, runner, and model summary

5. Windows notes. Model exposure tries an unprivileged same-volume hard link first, then a symbolic link (Developer Mode or Administrator), then an owned copy. Before copying, omm checks destination free space and reports that the model now consumes additional bytes. File junctions do not apply because model targets are files, not directories.

Set OMM_HOME before installation and on later runs to put the model hub on another volume. Runners omm installs itself (KoboldCpp, text-generation-webui, AnythingLLM, Msty) go to OMM_HOME\apps and installer downloads to OMM_HOME\tmp, so nothing large lands on the system drive. (AnythingLLM still keeps its bundled Ollama and starter model, ~5 GB, under %APPDATA% - its installer offers no way to move that.)

[Environment]::SetEnvironmentVariable("OMM_HOME", "D:\omm", "User")
$env:OMM_HOME = "D:\omm"

Install native shell completion once, then restart the shell:

omm --install-completion powershell

Remove a Git-source installation while preserving downloaded models and settings:

irm https://omm.run/uninstall.ps1 | iex

Download that script and run it with -Purge to remove the model hub and settings too.

Runner note: on Windows x64 the checklist downloads the official AnythingLLM and Msty installers and runs them silently into OMM_HOME\apps (no winget package exists for either); on ARM Windows it prints their download page instead.

Detailed walkthrough: https://omm.run/install/windows

1. Open this app. Open Terminal (Applications → Utilities → Terminal), or any terminal emulator you already use. The command runs under sh/zsh/bash.

2. Requirements. Python 3.10+ and git. The installer uses Homebrew when either is missing: if Homebrew itself is not installed, it bootstraps Homebrew with Homebrew's official installer first (Homebrew requires a supported macOS and Apple's Xcode Command Line Tools). To require a pre-existing Homebrew installation instead of letting the installer bootstrap it, export OMM_AUTO_INSTALL_HOMEBREW=0 before running the command. pipx is bootstrapped with pip if missing (Homebrew and other PEP-668 "externally-managed" Pythons are handled with --break-system-packages), and omm is then installed as an isolated CLI via pipx. The installer also persists pipx's executable directory in ~/.zprofile, so a newly opened zsh finds omm automatically. The optional NVIDIA detector is installed only when nvidia-smi indicates an NVIDIA driver.

3. Run the installer.

curl -fsSL https://omm.run/install.sh | sh

4. After install. Open a new shell afterward so your PATH picks up omm, then run:

omm        # first run: setup wizard (hardware scan + engine checklist)
omm scan   # hardware, runner, and model summary

5. Homebrew Tap (alternative).

brew install omm-hippo/omm/omm

Upgrade or remove the formula with Homebrew. Removing the formula preserves downloaded models and settings under OMM_HOME:

brew upgrade omm-hippo/omm/omm
brew uninstall omm-hippo/omm/omm

After a tagged PyPI release passes its public install checks, GitHub notifies the Homebrew Tap. Homebrew intentionally applies its upstream release cooldown before opening an automated Formula update PR, so the Tap can temporarily lag behind PyPI. Use brew info omm-hippo/omm/omm to see the version currently provided by the Tap. omm update does not modify a Homebrew installation and instead prints the matching brew upgrade command.

6. macOS notes. Set OMM_HOME before installation and on later runs to put the model hub elsewhere:

export OMM_HOME=/mnt/models/omm

Install native shell completion once, then restart the shell:

omm --install-completion bash  # or zsh/fish

Remove a Git-source installation while preserving downloaded models and settings:

curl -fsSL https://omm.run/uninstall.sh | sh

Download that script and run it with --purge to remove the model hub and settings too.

Detailed walkthrough: https://omm.run/install/macos

1. Open this app. Any terminal emulator. The command runs under sh/bash.

2. Requirements. Python 3.10+. The installer bootstraps python3, python3-venv/equivalent, and git via whichever supported package manager is present — apt-get, dnf, yum, pacman, or apk — when the current user can install system packages; on an unsupported distribution, install Python 3.10+ and git yourself first or the installer exits with Python 3.10+ not found or git not found. Install git first. pipx is bootstrapped with pip if missing (PEP-668 "externally-managed" Pythons are handled with --break-system-packages), and omm is then installed as an isolated CLI via pipx. The optional NVIDIA detector is installed only when nvidia-smi indicates an NVIDIA driver.

3. Run the installer.

curl -fsSL https://omm.run/install.sh | sh

4. After install. Open a new shell afterward so your PATH picks up omm, then run:

omm        # first run: setup wizard (hardware scan + engine checklist)
omm scan   # hardware, runner, and model summary

5. Linux notes. Set OMM_HOME before installation and on later runs to put the model hub on another volume:

export OMM_HOME=/mnt/models/omm

Install native shell completion once, then restart the shell:

omm --install-completion bash  # or zsh/fish

Remove a Git-source installation while preserving downloaded models and settings:

curl -fsSL https://omm.run/uninstall.sh | sh

Download that script and run it with --purge to remove the model hub and settings too.

Runner note: omm installs Jan on Linux through Flatpak, so flatpak must be present for that entry in the runner checklist.

Detailed walkthrough: https://omm.run/install/linux

Works on macOS, Linux, and Windows:

# macOS / Linux (Python 3.10+ and pip must already be installed)
python3 -m pip install omm-model

# Windows (Python 3.10+ and pip must already be installed)
py -m pip install omm-model

This does not go through the signed-commit verification described below; it relies on PyPI's own account security and TLS, the same trust model as installing any other PyPI package. It is a package-manager path, not a zero-prerequisite installer: install Python and pip first on a clean computer.

For an isolated command-line installation, pipx is recommended:

# If pipx is not installed yet, install it with the same Python first.
python3 -m pip install --user pipx
python3 -m pipx ensurepath
python3 -m pipx install omm-model

On Windows, use py -m pip, py -m pipx, and py -m pipx ensurepath instead. If the operating system marks Python as externally managed, install pipx from the operating system package manager or use the Git-source installer above.

The distribution name is omm-model; the installed command and Python import remain omm. Upgrade and remove it with the same tool that installed it:

# macOS / Linux
python3 -m pip install --upgrade omm-model
python3 -m pip uninstall omm-model

# Windows
py -m pip install --upgrade omm-model
py -m pip uninstall omm-model

# Or, for pipx:
pipx upgrade omm-model
pipx uninstall omm-model

Both commands preserve downloaded models and settings under OMM_HOME.

The public npm launcher installs a platform-specific standalone omm binary, so Python is not required for this path. It requires Node.js 22.14 or newer:

npm install --global @omm-hippo/omm
omm --version

Published npm binaries cover macOS on Apple Silicon and Intel, glibc-based Linux on ARM64 and x86_64, and Windows x64. Other operating-system, architecture, and libc combinations should use the Python-based installer or PyPI path instead.

Upgrade or remove the npm installation with:

npm update --global @omm-hippo/omm
npm uninstall --global @omm-hippo/omm

The npm launcher verifies the selected native package name, version, platform metadata, and binary containment before executing it. Release CI separately checks package integrity and registry signatures. Removal preserves models and settings under OMM_HOME.

Match the message you see, not the step you think you are on.

The Windows x64 portable ZIP and checksum are published as CI-verified GitHub Release assets. The community Winget submission is not yet merged, so there is no supported winget install command to document at this time.

omm is tested in CI on Windows, macOS, and Linux. Python-based installation paths require Python 3.10 or newer; the npm path requires Node.js 22.14 or newer and one of the published native targets listed above. Windows 10 22H2/11 is the supported Windows baseline because that matches Ollama's native Windows requirements. Hardware scan, install, linking, benchmark, update, and contribution flows are cross-platform. Benchmarking selects an available local Ollama or LM Studio runtime; runner- and model-specific capabilities can still differ.

Both installers clone to a versioned staging directory, verify the signed commit against a bootstrap trust anchor, and only then switch pipx to it. Do not replace this with an unverified git clone plus pipx install if commit authenticity matters.

omm update updates only a canonical OMM Git-source installation. For a package-managed installation, it leaves files unchanged and prints the matching package-manager upgrade command. The Git-only beta channel is likewise unavailable to package-managed installations. Git-source updates parse the new release's declared dependencies, honor python_version markers, and refresh the pipx environment when an applicable dependency is missing or older than the declared minimum.

Additional package-manager commands are added here only after their public registry path has been installed and verified.

The first bare omm run on a fresh install (or omm setup any time after) shows a hardware summary and a checklist of local AI runners. Checking one that omm knows how to install runs its official installer with live progress in the terminal; checking one it doesn't yet automate on your platform prints a link instead. Automation coverage today:

Every currently-installed runner is also listed (marked as already installed, not selectable) rather than hidden, so the checklist always reflects what omm actually detects on the machine.

The model hub and omm state default to ~/.omm. Set OMM_HOME before installation and on later runs to put them on another volume (see the snippet in your OS section above).

Ollama's own model location follows OLLAMA_MODELS. LM Studio follows its home pointer; set OMM_LMSTUDIO_MODELS_DIR when LM Studio uses a custom directory that omm cannot discover automatically.

Each OS section above carries the completion and uninstall commands for that shell. For a PyPI installation, use python -m pip uninstall omm-model; for pipx, use pipx uninstall omm-model. Both commands preserve downloaded models and settings under OMM_HOME.

Purge (-Purge on PowerShell, --purge on sh) removes only known omm-owned paths and leaves unrelated files in a custom OMM_HOME untouched. Installers mark custom homes so uninstallers can refuse ambiguous or unsafe locations; shell profiles are never rewritten during uninstall.

omm setup  # First-run setup wizard: hardware scan + engine checklist (re-runnable any time)
omm engine install [ENGINE]  # Install one supported local runner, or choose interactively
omm scan [--json]  # Print a hardware, runner, and model summary (RAM, VRAM, OS)
omm doctor [--json]  # Read-only diagnostics for the installation and Ollama reachability/links
omm recommend [--json]  # Rank compatible models, mark installed ones, and offer a new one to install
omm tune <name> [--json]  # Recommend context, GPU offload, threads, and batch size
omm search <query> [--json] [--skip-unfit] [--skip-ms] [--limit N] [--provider curated|huggingface|modelscope]  # Search curated, Hugging Face, and ModelScope sources
omm help [command]  # Show help, same as --help

omm recommend keeps compatible models visible when they are already present, marks whether OMM or another supported local runner owns them, and never reinstalls a selected installed model. In JSON output, installed, managed_by_omm, installed_engines, and installation_match expose the same state. installation_match distinguishes an exact artifact from the same model-and-parameter-size identity in a manifest-based runner, whose local quantization may differ. With --yes, the highest-ranked model that is not installed is selected; if every displayed recommendation is already present, the command exits without downloading.

omm install <name> [--skip-unfit] [--upload/--no-upload] [--force] [--verify-runtime|--no-verify-runtime]  # Download, link, and optionally verify a model
omm fit <name>  # Memory card: does this model (installed or not) fit next to what is running right now?
omm run [name] [--engine NAME]  # Chat with an installed model: Ollama in the terminal, KoboldCpp/text-generation-webui with the model loaded, GUI apps opened
omm import [directory] [--yes]  # Adopt GGUF files found across supported runners (and an optional directory) into the hub
omm uninstall <name> [--dry-run]  # Uninstall a model and clean up its symlinks/manifests (alias: rm)
omm uninstall all [--yes] [--dry-run]  # Uninstall every model installed via omm
omm list [--json] [--engine NAME]  # Show models installed via omm and their linked status (alias: ls)
omm info <name> [--json]  # Show a model's name, version, size, and linked-program run commands
omm upgrade <name> [--dry-run]  # Refresh a model against its source if it has changed since install (alias: up)
omm upgrade [--yes] [--dry-run]  # Check every installed model for updates
omm link [--engine NAME]  # Re-verify and repair installed-model links across supported runners
omm link <directory>  # Reuse central GGUF files; Windows warns if a real copy is required
omm autoremove  # Clean up broken symlinks in AI runner model directories
omm cleanup  # Clean up orphaned partial/incomplete downloads

install, uninstall, info, and upgrade accept either a model name/reference or the numeric index shown by the last omm search or omm list run in that terminal. search/install mark models predicted not to run on this machine's hardware in red.

omm install --skip-unfit is a scripting-friendly skip, not a successful installation: it prints Skipped and leaves the model hub unchanged. If an uninstall cannot remove the managed model file, OMM exits with status 1 and keeps the registry and any still-live links so the same command can be retried.

omm verify <name> [--engine ollama|lmstudio] [--yes] [--keep-loaded]  # Prove local load + generation works
omm benchmark <name>... [--output PATH]  # Local quality + speed evidence for selected installed models
omm benchmark all [--output PATH]  # Benchmark every installed model in the selected runtime
omm contribute [--yes]  # Repeatedly install/benchmark/upload hardware-fit models to grow the dataset

omm verify checks more than a link: it asks before starting a stopped local runtime or loading an unloaded model, sends one short deterministic prompt, requires a non-empty answer, and stops or unloads only what OMM started for the check. It never deletes the model or stores the generated answer. Use --yes for unattended confirmation and --keep-loaded to preserve a model OMM loaded for the check. LM Studio API authentication reads LM_API_TOKEN from the process environment and never writes it to config.json. Compatibility status is stored locally in models.json and is shown by omm info.

omm update  # Update a canonical OMM Git-source install; package installs print their manager command
omm setting  # Interactive menu for telemetry, upload policy, error reports, version, theme, calibration, and catalog trust
omm setting version [--stable|--beta]  # Show or switch the update channel `omm update` pulls from
omm setting telemetry --endpoint <url>  # Configure where benchmark telemetry is sent
omm setting upload --enable|--disable|--ask  # Configure the benchmark-upload send policy
omm setting error-reports --enable|--disable|--ask  # Configure the opt-in crash/error-report send policy
omm setting memory-guard --policy ask|block|observe  # Protect local runtime loads from live memory pressure
omm setting theme [--set NAME]  # Show or change omm's output color theme
omm setting calibrate <name>  # Locally correct predicted speed with an installed Ollama model
omm setting catalog-trust --manifest-url <url> --public-key <key>  # Require signed recommendation downloads
omm setting catalog-status  # Show signed recommendation data and rollback snapshots
omm setting catalog-rollback  # Restore the most recent different recommendation snapshot

All errors, warnings, and confirmation prompts print to stderr. For search, list, info, tune, scan, doctor, and recommend, --json makes stdout a single structured document that is safe to pipe (for example, omm list --json | jq .). benchmark --json appends its JSON report after the human-readable evidence summary, so treat the saved --output file as the machine-readable artifact instead of piping the complete stdout stream.

For commands that document --yes/-y, pass it to skip their confirmation prompts, or use the command-specific flag (install --skip-unfit, install --upload/--no-upload). Benchmark upload consent remains governed by omm setting upload; for non-interactive benchmarks, start the selected local runtime before invoking the command.

These common flags are available on most top-level commands. Where supported, they may be accepted before or after the subcommand name (omm --json search foo and omm search foo --json are equivalent). Use omm COMMAND --help for the exact flags and placement of a specific command:

• --json — structured output, where supported (see above)
• --yes / -y — skip confirmation prompts
• --quiet / -q — suppress progress bars and background status/hint lines (e.g. download progress, "Verifying checksum...", scan's "Run: omm link" nudge); errors, warnings, and the result of what you asked for still print
• --no-color — disable ANSI colors on omm's own console output and its download progress bar; the NO_COLOR environment variable does the same

Commands using the shared flag wrapper warn when --json or --yes has no effect. Exit codes are consistent across commands: 0 success, 1 failure, and 2 usage error (bad flag or argument).

rm, ls, and up are short aliases for uninstall, list, and upgrade.

Set OMM_HOME to store everything (models, config, catalog history) under a different directory instead of ~/.omm — useful when $HOME's filesystem doesn't have room for GGUF models, e.g. OMM_HOME=/mnt/data/omm omm contribute --yes.

omm contribute performs a 10 GiB startup free-space preflight. Before each download it separately budgets the central GGUF, a worst-case full runner copy, any required Windows cross-volume copies, and safety headroom. Each model evaluation prints a heartbeat every 30 seconds and is terminated after an absolute 10-minute deadline instead of hanging an unattended session indefinitely.

Localfit does not assume all installed memory belongs to the model. A live scan subtracts memory currently used by other applications, keeps at least 2 GB (or 10% of RAM) for the OS and newly opened apps, and applies total-memory caps. Recommendation fit and omm tune use this safe budget, so rerunning a command adapts after memory-heavy applications are opened or closed.

omm benchmark runs a versioned eight-item bilingual arithmetic smoke pack against models already installed in the selected Ollama or LM Studio runtime. It stores parsed answers, correctness, pinned model metadata, and fixed-length timings under OMM_HOME/evaluations/ (default ~/.omm/evaluations/); it stores no generated text. Opt-in telemetry sends a locally computed CPU chip score (and GPU chip score, when a GPU is present) plus architecture and core counts — never the raw CPU/GPU model name — so speed predictions can distinguish otherwise identical Linux x86_64 machines. Results are uploaded only after explicit opt-in. The pack is intentionally small and is not a leaderboard.

On Windows, Ollama is detected by its HTTP API first, so a freshly installed tray app works even before the current terminal receives the new PATH. When the daemon is stopped, omm also checks Ollama's documented %LOCALAPPDATA%\Programs\Ollama location. It only stops daemon processes it started itself. Before deleting a contribution model, omm requests an Ollama unload, waits for /api/ps to confirm handle release, and uses bounded retries for Windows file locks. Real-time antivirus can still delay a first load; the benchmark uses repeated samples and reports their median. Do not disable your antivirus for omm.

Benchmark results are never uploaded without explicit per-run consent or an always policy. New installations use the hosted proof-of-work gateway as the default destination, while existing local-only configurations stay local. To run the bundled FastAPI + SQLite collector instead:

pip install -e ".[server]"
export LOCALFIT_DB_PATH="$PWD/localfit.db"
export LOCALFIT_ADMIN_TOKEN="replace-with-a-long-random-token"
localfit-server

Explicitly configure the endpoint and opt in before uploading:

omm setting telemetry --endpoint http://127.0.0.1:8000/v1/benchmarks
omm setting upload --enable

Loopback ingestion needs no token. If the collector listens on a non-loopback interface, set the same LOCALFIT_INGEST_TOKEN on both the server and the omm client; remote ingestion fails closed when it is missing.

Training can consume the authenticated export directly:

export LOCALFIT_ADMIN_TOKEN="replace-with-a-long-random-token"
python scripts/train_model.py \
  --telemetry-url http://127.0.0.1:8000/v1/benchmarks/export

Firebase Realtime Database JSON endpoints remain supported. An official *.firebaseio.com or *.firebasedatabase.app .json URL can be read without an admin token; self-hosted raw export requires LOCALFIT_ADMIN_TOKEN. Exact duplicate events are ignored.

Automated retraining is fail-closed. Configure LOCALFIT_TELEMETRY_EXPORT_URL; configure LOCALFIT_ADMIN_TOKEN as well for a self-hosted export (it is optional for an official Firebase JSON URL). The scheduled job otherwise stops without changing the published artifact. It requires at least 100 distinct valid v6/v7 configurations with explicit runtime and CPU metadata (legacy rows do not satisfy this minimum), rejects datasets with more than 25% invalid rows, and reserves a deterministic 20% holdout. A 64-tree v4 candidate replaces the incumbent only when both holdout RMSLE and P90 absolute percentage error stay within the configured regression limits. Selection is evaluated on whole hardware/request contexts, so sibling model variants never leak across training and holdout sets. Publishing also requires at least three multi-model selection groups plus complete top-1, regret, balanced-fit, and false-positive evidence. Missing evidence fails the gate. The artifact records the complete candidate/baseline evaluation report.

The same gate can validate an exported local dataset without contacting the collector:

python scripts/train_model.py --offline \
  --telemetry-file benchmarks.jsonl \
  --quality-gate --minimum-real-configurations 100 \
  --baseline published/localfit-recommend-model.json \
  --output candidate.json --quality-report quality-report.json

Synthetic bootstrap training remains available for local development, but the scheduled publishing workflow never uses it as a substitute for missing real benchmark data.

omm setting catalog-trust --manifest-url <https-url> --public-key <base64-key> enables Ed25519 verification for future recommendation downloads. Existing artifacts are snapshotted before replacement and omm setting catalog-rollback restores the most recent different snapshot.

python -m venv .venv
source .venv/bin/activate  # Windows PowerShell: .venv\Scripts\Activate.ps1
python -m pip install -e ".[dev]" -r requirements-train.txt
python -m pytest -q

See CONTRIBUTING.md for platform-specific setup, scoped checks, signed-head requirements, and pull-request conventions.

Contributions are welcome. See CONTRIBUTING.md for setup, testing, and PR conventions, and CODE_OF_CONDUCT.md for community expectations. Report security issues per SECURITY.md rather than as a public issue.

Do not open a public issue for a suspected vulnerability. Follow the private reporting instructions and supported-version policy in SECURITY.md.

The omm source code is available under the MIT License. Third-party dependency notices are listed in THIRD_PARTY_NOTICES.md. Models and local runner applications installed or linked by omm keep their own upstream licenses and terms; the project MIT license does not replace them.