The Bundled Ollama Models

The Celestial Vault ships with a curated set of three Ollama models for first-launch auto-installation. The vault is pre-configured to use them — so the AI works without you running a single terminal command.

What Ships

The three bundled models:

All three are open-weight models you can run entirely on your own hardware. No cloud, no API keys, no token costs.

What "Bundled" Actually Means

The Celestial Vault doesn't ship the model weights themselves — those are too big to include in a downloadable zip. What it ships is:

1. Pre-configured plugin settings that point at these three model names
2. An auto-installation flow that runs on first launch, calling Ollama to pull each model for you
3. A model-setup modal that walks you through installation if Ollama isn't yet installed on your machine

So when you open the Celestial Vault for the first time and have already installed Ollama (see Installing Ollama), the welcome flow can fire ollama pull phi4-mini, ollama pull phi4, ollama pull deepseek-r1:1.5b for you in the background.

What If You Already Have Ollama Installed?

If you've used Ollama before — say you'd already pulled phi4 from a previous Time Garden Eternal install — Celestial detects this and skips re-pulling models you already have. No duplicate downloads, no wasted disk space.

Per-Feature Model Assignment

The bundled stack is designed so you can use different models for different operations:

The general pattern: fast model for short tasks, smart model for synthesis. You can change any of these in Obsidian Settings → Time Garden Plugin → AI Models.

(Full breakdown: Choosing Your AI Models.)

Why These Three?

Why phi4-mini as the default fast model

Microsoft's phi4-mini is the rare combination of:

• Small enough to run on most modern laptops without melting them
• Fast enough that daily AI buttons feel instant
• Surprisingly capable for short structured tasks (alias generation, ratings)

It's the workhorse. Most users will use it for 80% of operations.

Why phi4 for synthesis

For longer-form synthesis (monthly summaries, Year-in-Review), phi4-mini starts to feel thin. phi4 is the same family — same training lineage — but with the extra capacity to write a coherent paragraph instead of fragmented bullet points.

It's slower (phi4 runs at maybe a third of phi4-mini's speed on the same hardware), but the difference in monthly+ summaries is night and day.

Why deepseek-r1:1.5b

DeepSeek-R1 is a "reasoning" model — it thinks before it answers, which produces better output on questions that involve logic or reflection. The 1.5B variant is small enough to run alongside the others without memory issues.

It's an option, not the default. Try it for Q&A if phi4 feels too rigid.

Adding Other Models

The bundled three are not a hard cap. If there's another model you want to use:

1. Pull it: ollama pull <model-name> in your terminal
2. In Settings → Time Garden Plugin → AI Models, click Add Model, enter the name
3. Assign it to whichever operation in the dropdown

Anything Ollama can run, Celestial can use. (Full list at ollama.com/library.)

Removing Bundled Models

If your disk is tight and you want to free up space:

ollama rm phi4

removes the model from your machine. The next time you click an operation that's set to use phi4, Time Garden will throw an "model not found" error — at which point you can either re-pull it or assign a different model to that operation.

Resource Usage

While a model is running:

(Apple Silicon uses unified memory; Nvidia/AMD machines pull from GPU VRAM.)

If you're on an older laptop with 8 GB RAM, stick to phi4-mini and deepseek-r1:1.5b. phi4 will technically run but you'll feel it across the rest of your system.

When You Don't Want To Auto-Install

If you'd rather not have Celestial pull models on first launch — maybe you're on a metered connection, or you have a specific custom model you want to use instead — you can decline the auto-install during the Celestial welcome flow. The vault still opens; the AI just won't work until you set up models manually via Installing Ollama and Choosing Your AI Models.

Up Next

• Choosing Your AI Models — full per-feature configuration
• Customizing AI Prompts — tuning what each model is asked
• Dealing with useless AI answers — when output goes sideways