ollama-update

Check for Ollama model updates without pulling — by comparing local digests against the registry manifest. Only pulls when there's actually an update.

Why?

If you manage dozens of models, pulling each one just to check for updates wastes time and bandwidth. This script queries the registry first, so you get a fast dry-run view of what's outdated before committing to any downloads.

Features:

• Digest-based dry run — check what's outdated without pulling anything
• Custom Modelfile rebuild — detects when a base model updates and rebuilds your custom models
• VRAM lifecycle — unloads and reloads running models after update, preserving keep_alive timers
• Pinned models — skip specific models you don't want touched
• Config file — persistent settings without editing the script
• Remote endpoint — manage models on any Ollama instance
• JSON output — machine-readable results for scripting/CI
• CI-friendly exit codes — 0 all up to date, 1 updates available, 2 usage error

Quick start

# Check for updates (dry run — nothing is pulled)
ollama-update

# Actually pull updates and rebuild custom Modelfiles
ollama-update --update

# Pin models you don't want updated
ollama-update --pin llama3:70b,mistral:7b

# Check a remote Ollama instance
ollama-update --endpoint http://gpu-server:11434

# JSON output for scripting
ollama-update --json

Options

Flag	Description	Default
--update	Pull updates and rebuild custom Modelfiles	dry run
--json	Output results as JSON	off
--quiet	Only show UPDATE and SKIP lines	off
--defer-reload	Skip reloading updated models in VRAM	off
--config <path>	Config file path	~/.config/ollama-update/config
--modelfile-dir <path>	Custom Modelfile directory	~/.config/ollama-update/modelfiles
--endpoint <url>	Ollama API endpoint	http://localhost:11434
--pin <models>	Comma-separated models to skip	(none)
--jobs <N>	Parallel digest check jobs	unlimited (0)
-v, --version	Print version and exit	—

Exit codes

Code	Meaning
0	All models up to date (or updates applied with --update)
1	Updates available (dry run)
2	Error (bad args, endpoint unreachable, check failures)

Install

# Homebrew
brew tap bgupta/tap
brew install ollama-update

# Clone and symlink
git clone https://github.com/bgupta/ollama-update.git
ln -s "$(pwd)/ollama-update/ollama-update" ~/bin/ollama-update

# Or just download the script
curl -o ~/bin/ollama-update https://raw.githubusercontent.com/bgupta/ollama-update/main/ollama-update
chmod +x ~/bin/ollama-update

Update

# Homebrew
brew upgrade ollama-update

# Git clone
cd ollama-update && git pull

# Curl
# Re-run the curl command above

Requirements: bash 4.0+, curl, jq, ollama

Platform support

Platform	Status	Notes
Linux	Fully supported	Works with system Bash
macOS (Homebrew)	Fully supported	brew install bash (Bash 4.0+ required)
macOS (stock)	Not supported	Ships Bash 3.2 -- install via Homebrew
WSL / Windows	Fully supported	Works with system Bash

Configuration

Settings are loaded from ~/.config/ollama-update/config (bash key=value format):

# ~/.config/ollama-update/config
MODELFILE_DIR=$HOME/my-modelfiles
OLLAMA_ENDPOINT=http://gpu-server:11434
PINNED_MODELS=llama3:70b,codellama:34b

CLI flags override config file values. Config file overrides defaults.

Setting	CLI Flag	Config Key	Default
Config file	--config <path>	—	~/.config/ollama-update/config
Modelfile dir	--modelfile-dir <path>	MODELFILE_DIR	~/.config/ollama-update/modelfiles
Ollama endpoint	--endpoint <url>	OLLAMA_ENDPOINT	http://localhost:11434
Pinned models	--pin <model>[,...]	PINNED_MODELS	(empty)

Custom Modelfiles

If you use ollama create with custom Modelfiles (system prompts, parameters, etc.), put them in your modelfile directory as Modelfile-<name>, where <name> matches the installed model name (the :latest tag is implied):

~/.config/ollama-update/modelfiles/
  Modelfile-my-assistant     # FROM llama3:8b + custom system prompt
  Modelfile-coding-helper    # FROM codellama:7b + custom params

When a base model gets an upstream update, the script detects the digest mismatch and rebuilds your custom model automatically.

Output

Checking 24 models against registry (unlimited parallel)...

OK    llama3:8b
OK    mistral:7b
UPDATE codellama:7b
PIN   llama3:70b
SKIP  my-local-only-model:latest (not in registry)
OK    my-assistant (custom: base llama3:8b up to date)

Done: 3 ok, 1 update, 1 skip, 1 pin (2.4s)
Run with --update to pull updates and rebuild custom Modelfiles.

VRAM reload

When --update is used (without --defer-reload), the script:

1. Snapshots which models are currently loaded in VRAM and their keep_alive expiry
2. After updating, unloads and reloads affected models
3. Preserves the original keep_alive duration so they don't expire early

Use --defer-reload to skip this — updated models will cold-load on next use instead.

How it works

1. Discovers installed models via the Ollama HTTP API (/api/tags)
2. For each model, extracts the local weight blob digest via /api/show
3. Fetches the registry manifest from registry.ollama.ai and extracts the remote digest
4. Compares digests — if they differ, the model has an update
5. For custom Modelfile-based models, checks both the base model's registry digest and whether the custom model's blob matches its base
6. With --update: pulls changed models, rebuilds custom Modelfiles, reloads VRAM

The registry is always registry.ollama.ai regardless of your local endpoint setting — the endpoint only controls where ollama pull and VRAM operations target.

Contributing

Feature requests and bug reports are welcome — please open an issue.

License

MIT