omm

omm recommend

Rank models by a predictor trained on real install telemetry — falling back to static rules when the trained model can't be fetched — and offer to install the top pick.

01 / 06

Overview

Reach for recommend when you don't already know what to install: it scans this machine, ranks candidates by predicted speed and how much of the safe memory budget they'd use, and — outside --json — walks you through picking one and installs it directly. --json is read-only: it prints the ranked list and installs nothing, which is what makes it safe to script. --yes skips the picker and installs whatever ranked first.

02 / 06

Options

Every flag this command accepts, and what it defaults to when you leave it out.

  • --profilededicated | balanced | minimalDefault: prompted; balanced with --yes/--json

    Choose how much of this machine the model may claim: dedicated, balanced, or minimal. Interactive runs ask; --yes and --json default to balanced.

  • --json—Default: off

    Print the ranked candidates as JSON and install nothing.

  • --yes—Default: off

    Skip the interactive picker and install the top-ranked candidate immediately.

03 / 06

Examples

From a plain search to something you'd put in a script.

Interactive — ranks candidates, then walks you through picking one to install.

$ omm recommend

Prefer the smallest memory footprint so more of the machine remains available for other work.

$ omm recommend --profile minimal

Read-only — prints the ranked list, installs nothing.

$ omm recommend --json

Non-interactive — installs the top-ranked candidate without asking.

$ omm recommend --yes

04 / 06

A real run

Real omm recommend run, 2026-08-25, driven end to end through a real terminal against a mid-range PC (Intel Core Ultra 7 155H, 15.5 GB RAM, Intel Arc — the same machine design/FACTS.md's install guides use) instead of this session's own laptop, so the ranked list reflects hardware someone would actually run this on. Only the hardware reading was substituted; the ranking, the ten real candidates fetched live from GitHub, the arrow-key walk down to mistral 7b instruct v0.2 (a real candidate in that same ranked run), and the detail card after picking it are all genuinely computed and real, including its repository field — the same TheBloke/Mistral-7B-Instruct-v0.2-GGUF this site's own install demo already verifies. What selecting it actually downloads is shown too, reusing that same already-verified install (real byte count, real link summary for Ollama/LM Studio/Jan) rather than triggering a second real 4.4 GB download just for this page.

06 / 06

If something goes wrong

Every message below is one this command actually prints. Find yours, read why it happened, then do the last line.

  1. No model is predicted to run on this hardware.
    why
    Every candidate the trained predictor ranked came back under the minimum usable speed for this hardware.
    what to do
    This machine likely needs a smaller model than anything currently in the trained catalog — try omm search for something specifically small, e.g. a 1-3B model.
    source
    src/omm/cli.py:2916
  2. No model in the current rules fits this hardware.
    why
    The trained model wasn't available, so recommend fell back to the static rules — and even those found nothing that fits the memory this machine has free.
    what to do
    Close other applications to free memory and try again, or search for a smaller model directly.
    source
    src/omm/cli.py:2952

Still stuck? Open an issue with the exact message you saw.

All commands