Preface

This post reflects strong personal opinions. If you feel uncomfortable while reading, please close it immediately. This article is for personal learning records only. You are welcome to repost or share it under the license terms — please respect the copyright and keep the original link. Thank you for your understanding and cooperation. If you find this site helpful, consider subscribing via RSS. Thanks for your support!

1. What is local free translation

A large language model (LLM) does more than chat and code — translation is one of its strongest skills. But most people’s first thought is ChatGPT or a translation API: online, account-gated, paid, and your content leaves the country.

Your own Mac can run an open-source model and translate Chinese Markdown into English fully offline — free, no upload, works without internet. This post walks you through the complete setup from zero.

The three pieces:

Ollama: a local model runtime (runs the model; think Docker for containers)
Qwen2.5: Alibaba’s open-source model, strong in both Chinese and English; 7B runs smoothly on a Mac
Your browser: calls the local model through a tool page; no command line needed for daily use

graph LR
  A[You] -->|Chinese Markdown| B[Browser tool / CLI]
  B -->|fetch localhost:11434| C[Ollama]
  C --> D[Qwen2.5:7b local inference]
  D -->|English Markdown| B

2. Prerequisites (read before installing)

Hardware requirements

Item	Minimum	Recommended	Notes
Mac chip	Apple Silicon (M1+)	Any M1/M2/M3/M4	Intel works but is slow; Apple Silicon has MLX acceleration
RAM	8 GB	16 GB	8 GB can only run 3B; 16 GB handles 7B easily; 32 GB for 14B
Free disk space	15 GB	20 GB+	Model weights ~5–10 GB, Jekyll + Ollama ~2 GB
macOS version	macOS 12 (Monterey)	macOS 13 / 14 / 15	Ollama requires macOS 12+

How to check your RAM? Apple menu → About This Mac → the “Memory” row. Same place for chip type.

What you need to know beforehand

Comfortable with the Terminal app (all commands are copy-paste, no compilation)
Know the difference between “paste into terminal” and “press Enter”
No English requirement — every command is copy-paste

3. Install from scratch (0 to 1)

One-shot script: don’t want to type step by step? The script at tools/setup-local-llm-translator.sh does all of this in one go. Run it:
1 bash ~/Documents/sunyazhou/tools/setup-local-llm-translator.sh
It only runs on macOS (checks OS up front and exits on anything else). It skips any step that’s already done (Homebrew, Ollama, models). Check section “4. Verify your setup” after running to confirm readiness.

Step 1: Install Homebrew (skip if already installed)

Homebrew is macOS’s package manager — think App Store for the command line. Used to install Ollama and for future upgrades.

Check if already installed:

brew --version

If you see Homebrew 4.x.x you’re done — skip to Step 2.

Install if not:

  
/bin/bash -c "$(curl -fsSL https://raw.githubusercontent.com/Homebrew/install/HEAD/install.sh)"

It will ask for your Mac login password twice (normal, no echo) and once for Enter to confirm the install path.

Apple Silicon (M1 and later) only: after installing, Homebrew will prompt you to add its path to your shell config. Copy and run those two lines, otherwise brew won’t be found:
1 2 (echo; echo 'eval "$(/opt/homebrew/bin/brew shellenv)"') >> ~/.zprofile eval "$(/opt/homebrew/bin/brew shellenv)"
Intel Mac: path is /usr/local/bin/brew; this step is not needed.

Step 2: Install Ollama

Ollama is the “model runner”. It is analogous to Docker — you install the runtime, but no images/models come with it.

brew install ollama

Expected output (excerpt, last few lines):

🍺  Congratulations! Ollama is installed.
   Run `ollama serve` to start the server.

On macOS, Ollama usually auto-starts the service after installation, but not always. You can skip the “manually start serve” step and run ollama list first — if the service isn’t up yet, come back and start it manually.

Verify Ollama is installed:

ollama --version

Output like ollama version 0.x.x is good.

Step 3: Start the Ollama service

Translation tools and the CLI both connect to localhost:11434, so Ollama must keep running in the background.

Option A: Foreground (debugging, see logs)

ollama serve

The terminal will show Ollama is running and hang — don’t close this window, it’s the service process.

Option B: Background (recommended; run once after install)

  
nohup ollama serve > /tmp/ollama-serve.log 2>&1 &
echo "Ollama started in background, PID: $!"

Returns immediately to the prompt; closing the Terminal doesn’t stop the service. Logs are at /tmp/ollama-serve.log — check with cat /tmp/ollama-serve.log.

Option C: macOS LaunchAgent (auto-start on boot, recommended for daily use)

Make Ollama start automatically every time you log in — never worry about ollama serve again:

  
# Create the plist file (run this once)
mkdir -p ~/Library/LaunchAgents
cat > ~/Library/LaunchAgents/com.ollama.ollama.plist << 'EOF'
<?xml version="1.0" encoding="UTF-8"?>
<!DOCTYPE plist PUBLIC "-//Apple//DTD PLIST 1.0//EN" "http://www.apple.com/DTDs/PropertyList-1.0.dtd">
<plist version="1.0">
<dict>
    <key>Label</key><string>com.ollama.ollama</string>
    <key>ProgramArguments</key><array>
        <string>/opt/homebrew/bin/ollama</string><string>serve</string>
    </array>
    <key>RunAtLoad</key><true/>
    <key>KeepAlive</key><true/>
</dict>
</plist>
EOF

# Load and start
launchctl load ~/Library/LaunchAgents/com.ollama.ollama.plist
echo "Set to auto-start. Verify with: launchctl list | grep ollama"

Intel Mac: replace /opt/homebrew/bin/ollama with /usr/local/bin/ollama.

Verify the service is up:

curl -s http://localhost:11434/api/tags

Returns {"models":[]} or similar JSON — service is live. If you get Connection refused, go back to Option A/B/C to start it.

Step 4: Pull the model to your machine

Ollama is installed, but the model itself must be downloaded separately. ollama pull is like brew install for model weight files, stored in ~/.ollama/models/.

Model	Size	Best for	RAM needed
`qwen2.5:3b`	~2 GB	Fast first drafts, long posts, low-end machines	8 GB+
`qwen2.5:7b`	~4.7 GB	Default: best quality/speed balance	16 GB+
`qwen2.5:14b`	~9 GB	Highest quality, exact terminology required	32 GB+

ollama pull qwen2.5:7b   # recommended: best quality
ollama pull qwen2.5:3b   # lightweight: faster, good for rough drafts

Expected output (first-time pull, shows progress):

pulling manifest
pulling 8ec2f7aad86b... 100%  ████████████████████  4.7 GB
pulling 4c9d85b5d93f... 100%  ████████████████████    137 B
verifying sha256 digest
writing manifest
success

Expect 4–15 minutes depending on network speed. Works offline once downloaded — weights live in ~/.ollama/models/.

If the pull fails mid-way (network issue), just re-run ollama pull qwen2.5:7b. Ollama supports resume — it continues from where it left off, not from scratch.

Verify model is installed:

ollama list

Expected output:

NAME                 SIZE      MODIFIED
qwen2.5:7b           4.7 GB    2 minutes ago
qwen2.5:3b           1.9 GB    5 minutes ago

4. Verify your setup is ready

Run these two commands to confirm everything is green:

  
# Check 1: Ollama service online
curl -s http://localhost:11434/api/tags >/dev/null && echo "✓ Service is up" || echo "✗ Service not running — run ollama serve"

# Check 2: model installed
ollama list | grep -q "qwen2.5" && echo "✓ Models ready" || echo "✗ No model found — run ollama pull qwen2.5:7b"

Both ✓ — you’re done with setup. Time to translate.

5. How to use it

Option A: Command line (30-second check)

Translate text without opening a browser:

echo "Translate this to natural English: Using Codable to parse JSON is very convenient" | ollama run qwen2.5:7b

Expected output (a few seconds):

Using default招待模式。Switch modes within a session using /set mixture.
Parsing JSON with Codable is very convenient.

If you end up in interactive mode (>>> prompt), press Ctrl+D (or type /bye) to exit.

Option B: Blog visual tool page (recommended for writing posts)

Translating a whole Markdown post involves front matter, code blocks, and URLs — doing it all in the CLI is tedious. The tool page at /tools/md-translator/ automates this: paste Chinese Markdown, click once, code and links stay intact.

Step by step:

Step 1: Start the blog locally (required so the browser can reach your local Ollama)

  
cd ~/Documents/sunyazhou   # your blog directory
bundle install             # first time only; skip on subsequent runs
bundle exec jekyll s -l -o

If you see Could not find gem 'jekyll-polyglot', run gem install jekyll-polyglot or bundle install. If it hangs at Bundle complete!, open a new terminal tab and run bundle exec jekyll s there.

Step 2: Open the tool page

In your browser, navigate to: http://127.0.0.1:4000/tools/md-translator/

⚠️ Must use 127.0.0.1 or localhost. Do not open with file://. Avoid 0.0.0.0 — some machines route 0.0.0.0 traffic differently, and the browser won’t reach localhost:11434.

Step 3: Translate

Button	What it does
Translate	Translates the input area with the selected model
Load Sample	Fills in sample Chinese Markdown — click Translate to see the effect
Save Locally	Chrome / Edge: native macOS save dialog (recommended); Safari / Firefox: downloads to default folder
Copy	Copies output to clipboard
Clear	Clears both input and output

Step 4: Save to your blog

After translation:

Click “Save Locally” → Chrome shows the macOS native save dialog
Navigate to your blog’s _posts/en/ folder
Confirm the filename (the tool auto-names it using the front matter date + English slug)
Click “Save”

Example filename: 2026-08-27-local-llm-markdown-translator.md

The tool’s protection mechanisms (why feeding a raw file to the model would mangle code and URLs):

Front matter: only title / description are translated; date, tags, categories etc. are kept verbatim
Fenced code blocks ` ```swift `: locked with a placeholder, restored after — the model never touches code
Inline code `Codable`: same protection
Links / image URLs: only the visible text is translated; the address itself is untouched (even title attributes are preserved)
Glossary: terms entered in “Locked glossary” (comma-separated, e.g. KVO, KVC, Codable) are locked and never translated — keeps terms consistent across the whole file
Chunked by blank lines: long posts are split and translated concurrently; chunks with no Chinese are skipped; everything is losslessly stitched back together

Minimal example (verifying protection works):

Input (Chinese):

  
# 用 Codable 解析 JSON

Swift 里 `Codable` 很好用，参考 [文档](https://developer.apple.com/codable)。

```swift
struct User: Codable { let name: String }
```

Output (English, code and URL intact):

  
# Parse JSON with Codable

`Codable` is very useful in Swift. Check the [docs](https://developer.apple.com/codable).

```swift
struct User: Codable { let name: String }
```

This tool is for your own writing — not for site visitors. The live blog is pure static Jekyll with no backend; a visitor’s browser localhost:11434 points at their computer, not yours. Only you can use it while running jekyll serve locally.

6. Troubleshooting

Connection issues

Symptom	Cause	Fix
Translation says “can’t reach Ollama”	`ollama serve` not running	Re-run: `ollama serve` (new terminal tab; don’t close it)
`curl: (7) Failed to connect to localhost:11434`	Port not open / service down	`lsof -i :11434` to check; `ollama serve` to start
Can’t connect after sleep / reboot	Service was killed	Re-run `ollama serve` (models stay, no re-pull needed)
Tool page blank	Not opened via jekyll local address	Use `http://127.0.0.1:4000/tools/md-translator/`, not `file://`
Browser shows `0.0.0.0:4000`	Jekyll bound to all interfaces	Kill Jekyll and re-run `bundle exec jekyll s -H 127.0.0.1 -l -o`

Model issues

Symptom	Cause	Fix
“Model not found” error	Model not installed	`ollama list` to check; `ollama pull qwen2.5:7b`
`ollama pull` stuck at 0%	Network cannot reach ollama.com	Switch to a better network / wait; supports resume — re-run to continue
Pull failed mid-way	Network glitch	Re-run `ollama pull qwen2.5:7b`; it resumes from where it left off
Pull too slow	Slow network	Try late night or a better network; qwen2.5:3b is only 2 GB and much faster

Memory issues

Symptom	Cause	Fix
Mac gets hot / fans loud during translation	Normal (model inference is CPU/GPU intensive)	Expected; Mac thermal design handles this fine
Translation hangs then times out	Not enough RAM for 7B	Switch to 3B: `ollama pull qwen2.5:3b`; change model box to `qwen2.5:3b`
`Killed: 9` or “Out of memory”	8 GB Mac trying to run 14B	Downgrade to 3B or 7B

Quality issues

Symptom	Cause	Fix
Translation modified code block content	Placeholder protection bug	Report it; use “Load Sample” first to confirm code blocks stay intact
Translation changed a URL	Chinese params in URL got translated	Tool should handle this; report if you see it
Terminology inconsistent across file	Model didn’t recognize a term	Add it to the “Locked glossary” field (comma-separated) and re-translate
Output is garbled	Model inference error	Switch to 3B or re-pull: `ollama rm qwen2.5:7b && ollama pull qwen2.5:7b`

Other

Symptom	Cause	Fix
`ollama: command not found`	Homebrew PATH not configured	Add to `~/.zshrc`: `eval "$(/opt/homebrew/bin/brew shellenv)"` then run it
`ollama serve` says port in use	Another process on port 11434	`lsof -i :11434` to find PID; `kill -9 <PID>` to stop it
Jekyll says `Could not find gem`	Missing Ruby gem	`bundle install`
`bundle install` permission error	Blog directory ownership issue	`sudo chown -R $(whoami) ~/Documents/sunyazhou`

7. Swap models

ollama pull qwen2.5:14b   # highest quality, but 16 GB RAM gets tight; 32 GB recommended

Change the model name in the tool page’s “Model” input box to qwen2.5:14b (or any pulled model name) — no restart of Ollama or Jekyll needed, just refresh and try.

8. Long-term maintenance

Update Ollama

brew upgrade ollama

Update a model (re-pull latest version)

ollama pull qwen2.5:7b   # downloads and replaces with the newest version

List installed models

ollama list

Remove an unused model

ollama rm qwen2.5:3b   # frees ~2 GB

Full uninstall (clean slate)

  
# Stop LaunchAgent (if configured)
launchctl unload ~/Library/LaunchAgents/com.ollama.ollama.plist
rm ~/Library/LaunchAgents/com.ollama.ollama.plist

# Uninstall Ollama
brew uninstall ollama

# Delete all model weights (~10 GB; asks for confirmation)
rm -rf ~/.ollama

# Remove Homebrew (optional)
/bin/bash -c "$(curl -fsSL https://raw.githubusercontent.com/Homebrew/install/HEAD/uninstall.sh)"

9. Command reference

Scenario	Command
Install Homebrew	`/bin/bash -c "$(curl -fsSL https://raw.githubusercontent.com/Homebrew/install/HEAD/install.sh)"`
Install Ollama	`brew install ollama`
Start service (foreground)	`ollama serve`
Start service (background)	`nohup ollama serve > /tmp/ollama-serve.log 2>&1 &`
Check service is up	`curl -s http://localhost:11434/api/tags`
Pull 7B model	`ollama pull qwen2.5:7b`
Pull 3B model	`ollama pull qwen2.5:3b`
List installed models	`ollama list`
Delete a model	`ollama rm qwen2.5:7b`
Translate one line via CLI	`echo "sentence to translate" \\| ollama run qwen2.5:7b`
Start blog	`cd ~/Documents/sunyazhou && bundle exec jekyll s -l -o`
Open tool page	`http://127.0.0.1:4000/tools/md-translator/`
Update Ollama	`brew upgrade ollama`

10. Tips and wrap-up

Default to 7B: best balance of quality and speed; 16 GB Mac handles it without breaking a sweat.
Keep a glossary: put project-specific terms in the glossary field for consistent translation throughout the file.
Privacy first: fully local — safe to paste unpublished content.
Free & offline forever: install once, zero ongoing cost.
LaunchAgent saves the day: configure it once and Ollama starts automatically on every login — never hunt for ollama serve again.

From installing Homebrew and Ollama, pulling Qwen2.5, to translating a full blog post via the tool page — not a cent spent, nothing leaving your machine. That’s the 0 to 1 of local free translation.

Offline Blog Translation with a Local LLM: Ollama + Qwen2.5

Preface

1. What is local free translation

2. Prerequisites (read before installing)

Hardware requirements

What you need to know beforehand

3. Install from scratch (0 to 1)

Step 1: Install Homebrew (skip if already installed)

Step 2: Install Ollama

Step 3: Start the Ollama service

Step 4: Pull the model to your machine

4. Verify your setup is ready

5. How to use it

Option A: Command line (30-second check)

Option B: Blog visual tool page (recommended for writing posts)

6. Troubleshooting

Connection issues

Model issues

Memory issues

Quality issues

Other

7. Swap models

8. Long-term maintenance

Update Ollama

Update a model (re-pull latest version)

List installed models

Remove an unused model

Full uninstall (clean slate)

9. Command reference

10. Tips and wrap-up

Trending Tags

Offline Blog Translation with a Local LLM: Ollama + Qwen2.5

Preface

1. What is local free translation

2. Prerequisites (read before installing)

Hardware requirements

What you need to know beforehand

3. Install from scratch (0 to 1)

Step 1: Install Homebrew (skip if already installed)

Step 2: Install Ollama

Step 3: Start the Ollama service

Step 4: Pull the model to your machine

4. Verify your setup is ready

5. How to use it

Option A: Command line (30-second check)

Option B: Blog visual tool page (recommended for writing posts)

6. Troubleshooting

Connection issues

Model issues

Memory issues

Quality issues

Other

7. Swap models

8. Long-term maintenance

Update Ollama

Update a model (re-pull latest version)

List installed models

Remove an unused model

Full uninstall (clean slate)

9. Command reference

10. Tips and wrap-up

Keep Reading

macOS Mouse Software MOS

Avoid Using Spacer in SwiftUI Stacks

Crash Types in iOS

Trending Tags