Ollama 教學：三個指令在 MacBook M1 跑本地 AI，資料完全不出去Ollama Quickstart: Three Commands to Run Local AI on MacBook M1, Data Never Leaves

So I set up a local LLM instead. Here's how it went.

Requirements

8GB will work but you'll hit swap — slow output, loud fan. 16GB+ is the smooth experience.

Three commands

1. Install Ollama

Download the macOS installer from ollama.com, drag to Applications, launch. A llama icon appears in the menu bar — that's the background service running.

Confirm it's installed:

ollama --version

If you see a version number, you're good.

2. Pull the model

ollama pull gemma4:e2b

This downloads ~7.2GB of model weights to your machine. Takes 10–20 minutes depending on your connection.

Why gemma4:e2b? It's Google's Gemma 4 Efficient 2B — roughly 2 billion parameters, fits comfortably in 16GB RAM. Traditional Chinese quality is noticeably better than the previous generation.

📌 Why not DeepSeek?

DeepSeek is developed by a Chinese company. Several versions have data handling policies that conflict with Taiwan's Personal Data Protection Act requirements. Resonance Stack only recommends models with clear data sovereignty from non-Chinese teams. Blocked list: Qwen, Yi, DeepSeek, ChatGLM.

3. Start chatting

ollama run gemma4:e2b

Takes 10–15 seconds to load, then you're in.

The gotcha: it defaults to English

This is the thing other tutorials skip.

gemma4:e2b responds in English by default. Even if you write your question in Chinese, the first reply will likely come back in English. The fix is a one-line prefix at the start of every prompt:

請務必使用繁體中文台灣用語回答。

(Translation: "Please respond in Traditional Chinese with Taiwanese vocabulary.")

Add that line, and it works correctly. This is the model's default behavior — nothing to do with Ollama itself. (I burned one prompt so you don't have to.)

The actual test: my rental contract

I pasted three clauses from Taiwan's government-standard rental contract (Ministry of the Interior, latest version April 2026) directly into the prompt:

請務必使用繁體中文台灣用語回答。

以下是台灣內政部公版住宅租賃定型化契約的三個條款：

第五條 押金：最高不得超過二個月租金之總額，
租期屆滿房東須返還押金。

第六條（電費）：以用電度數計費者，
每度電費不得超過台電帳單「當期每度平均電價」。

第九條 修繕：房屋損壞原則由房東負責修繕。
如房東遲不修繕，房客可催告，催告無效後得自行修繕並請求償還費用。

我下週要簽這份契約，身為房客有什麼要特別注意的？
請用白話說明，重點就好，不用長篇大論。

(The prompt stays in Chinese — that's what actually gets sent to Ollama.)

31 seconds. The model correctly explained all three clauses: deposit cap (two months max), electricity must follow Taiwan Power's actual rate, repair request procedure. Traditional Chinese was clean, legal terms were understood correctly. Added some emoji (🔑 💡) — readable but unmistakably AI-flavored. Quality: 4/5.

Follow-up 1: Is NT$5 per unit legal?

Clause 6 (electricity) prompted an obvious follow-up. Many landlords in Taiwan charge a flat NT$5 per kWh, but Taiwan Power's actual average rate is around NT$3.5. Is that legal?

請務必使用繁體中文台灣用語回答。

背景：台灣內政部公版住宅租賃定型化契約第六條規定：
「以用電度數計費者，每度電費不得超過台電帳單『當期每度平均電價』。」

問題：房東說每度電費收5元。台電帳單當期每度平均電價是3.5元。
房東這樣收合法嗎？假設每月用100度電，差多少錢？身為房客可以怎麼做？

36 seconds:

• Legal? No. NT$5 exceeds the NT$3.5 cap — violates Clause 6.
• How much extra? 100 units × NT$3.5 = NT$350 (legal max). Landlord charging NT$500. That's NT$150 overpayment per month — about US$4.50.
• What can I do? Three steps: collect Taiwan Power bills as evidence → send written notice citing Clause 6 → escalate to the housing authority if no response.

Correct math, correct legal citation, no fluff. 4/5.

Follow-up 2: My contract bans me from applying for housing subsidies. Legal?

This one gets more interesting. Some landlords add a clause like: "Tenant may not apply to any government rental subsidy or housing allowance program." Looks like a mutual agreement — but is it enforceable?

I tested it against Prohibited Clause No. 10 from the Ministry of the Interior's own contract template:

請務必使用繁體中文台灣用語回答。

台灣內政部公版住宅租賃定型化契約 不得記載事項第十點 規定：
「不得記載承租人不得申請租金補貼。」

情境：我在準備簽的契約裡，發現房東加了一條額外條款：
「承租人不得向政府申請任何形式之租金補貼或住宅津貼。」

問題：
1. 房東這條合法嗎？
2. 我如果簽了，這條有效嗎？
3. 台灣的租金補貼怎麼申請？申請成功的話，我的房東會受到什麼影響？

45 seconds. Worth unpacking each answer:

Q1 (Is it legal?): The model was hedgy — it circled around "contract freedom has limits" without clearly stating this clause directly violates the Ministry's prohibited clause list. The correct answer is more direct: Prohibited Clause No. 10 explicitly bans this. It's illegal, full stop.

Q2 (Enforceable if I sign?): This is where the model got it wrong. It said the clause is "formally valid." That's incorrect — under Taiwan's Consumer Protection Act Article 12, standard contract clauses that violate mandatory regulations are void. Even if you sign, the clause cannot be enforced against you.

Q3 (How to apply for subsidies?): Reasonable answer — apply through local government with income and residency documentation. Direction is correct.

Overall: 3/5. The model was too cautious on the legal questions, stopping short of clear conclusions. This is a real limitation of a 2B model on legal reasoning — it points you in the right direction but won't commit to a verdict. Use it to prepare your questions, not to reach your conclusions.

Running cost

MacBook M1 draws about 15–20W during inference. At Taiwan Power's current rates, running it 2 hours a day costs roughly NT$3–5 per month.

This isn't a Claude Pro replacement — the rental subsidy question showed the gap on legal reasoning. But for "help me understand this document" and "flag the suspicious clauses" tasks, three commands, seven minutes, and your data stays on your machine. That's a low bar to clear.

Memory footprint and inference speed

This was taken while Claude was also open — both apps running simultaneously. The numbers tell a clean story:

• ollama (gemma4:e2b loaded): 7.28 GB — roughly equal to the model file size, as expected
• Claude processes (Claude, claude, Claude Helper × 2): ~1 GB combined
• System overall: memory pressure graph stays green throughout, swap at just 844.3 MB

Two apps running, machine not sweating, fan not spinning. That's Apple Silicon's unified memory architecture doing what it's supposed to do.

Inference speed: the three experiments ran in 31s, 36s, and 45s — roughly 20–30 tokens per second. Slower than ChatGPT, but for "paste a document and ask a question" workflows it's plenty. You ask, wait thirty seconds, read the answer. That cadence doesn't interrupt a work session.

FAQ

Q: What's the main difference from ChatGPT? A: Your data doesn't leave your machine. Whatever you type, whatever you paste — stays on your computer. No server logs, no transmission fees, no terms of service edge cases.

Q: Is the Traditional Chinese quality good enough? A: For everyday document analysis (rental contracts, utility calculations), yes. For precise legal advice, still talk to a lawyer — AI is a research tool, not a verdict machine. Don't forget the Chinese prefix on every prompt or it'll answer in English.

Q: Can I run this on 8GB RAM? A: Yes, but you'll hit swap — noticeably slower output, fans might spin up. 16GB+ is the comfortable experience.

Verdict

Waveform: CLEAN_SIGNAL

Good fit if you:

• Have documents (contracts, agreements) you don't want to upload to the cloud
• Have a MacBook M-series with 16GB+ RAM
• Are already paying for too many AI subscriptions
• Don't mind a one-line prefix on every prompt

Not the right tool if you:

• Need precise legal conclusions (it hedges)
• Have 8GB RAM and care about speed
• Want zero Terminal interaction

For what it's worth: I still subscribe to both Claude Pro and Gemini Advanced — they each have strengths I'm not ready to give up. But gemma4:e2b gets more of my day-to-day usage than I expected, particularly for Chinese semantic analysis and annotation tasks that run repeatedly. No token quotas to think about — just run it as many times as you need. That whole workflow deserves its own post.

NOTE: gemma4:e2b's thinking mode outputs its reasoning process before answering, filling your Terminal with ANSI escape characters. It's not broken — just scroll past the noise. The real answer is at the end.

Visit Ollama →