Choosing Your AI Models

Time Garden Eternal lets you pick a different AI model for every operation. That's a lot of dials — but you don't have to use them. Defaults are good. This page is for when you want to make them better.

---

The Quick Recommendation

If you don't want to think about it:

• Lightest setup: phi4-mini for everything (fast, runs anywhere, ~2.5 GB)
• Recommended setup: phi4-mini for daily operations, phi4 for weekly+ (much smarter for long-form synthesis, ~9 GB)
• Power setup: phi4-mini for daily, phi4 for weekly/monthly, mistral:instruct for Q&A (~4 GB)

You can always start light, then upgrade later. Models can be added or removed at any time via ollama pull and ollama rm.

---

What Each Operation Uses

Time Garden lets you assign a model to each of the following:

Daily

Weekly

Monthly / Quarterly / Yearly

Chunking

(See Chunking Settings Explained for what chunking is.)

---

How To Change Models

1. Open Obsidian Settings (gear icon, bottom-left)
2. Click Time Garden Plugin in the left sidebar
3. Find the AI Models section
4. For each operation, pick a model from the dropdown

The dropdown only shows models you have installed locally. If a model you want isn't there, pull it via Ollama (the terminal command) or use the Add Model button in Time Garden's settings to register it.

---

Which Models Are Good For What?

Small / Fast (2–4 GB)

• phi4-mini — Microsoft. Excellent for short tasks like aliases and ratings. The default for a reason.
• deepseek-r1:1.5b — Tiny reasoning model. Sometimes hallucinates on summarization. Decent for simple tasks.

Medium (4–9 GB)

• mistral:instruct — Strong general-purpose, good at instructions, fast
• llama3.1:8b — Solid all-rounder
• phi4 — Microsoft's larger model. Excellent for summarization and Q&A. The recommended upgrade.

Larger (10+ GB)

• llama3.3:70b — Top-tier; needs serious hardware (32GB+ RAM)
• mixtral — Excellent for long-context tasks; large
• qwen2.5:32b — Strong all-rounder

---

A Strategy

1. Start with phi4-mini everywhere. Live with it for a week.
2. Notice what feels weakest — usually weekly+ summaries and Q&A.
3. Upgrade just those operations to phi4 or mistral:instruct.
4. Leave the daily-rating / daily-alias on phi4-mini — these are short tasks and don't benefit much from a heavier model.

That gets you 80% of the quality at maybe 30% of the runtime cost.

---

Adding Custom Models

Anything Ollama can pull, Time Garden can use. So if you find a model on ollama.com/library that excites you:

1. ollama pull <model-name> in your terminal
2. In Time Garden Plugin settings, click Add Model, enter the name
3. It now appears in every dropdown

There's also support for fine-tuned and quantized variants — anything Ollama can run will work.

---

One Important Caveat

---

Up Next

• Customizing AI Prompts — tune what each model is asked
• Chunking Settings Explained — control how much context each model sees
• Dealing with useless AI answers — when output goes sideways