ClariRec ships with a free local translation model that runs entirely on your Mac. After a one-time download of about 1.1 GB, translation works offline — no API key, no cloud account, no network required for the translation itself.
This page covers the local model specifically. For the full picture of all translation engines ClariRec supports, see Translate screenshot text. For translating without a screenshot, see Translate selected text.
What the local model is
The model is Tencent Hunyuan Machine Translation 2 (HY-MT2), a 1.8-billion-parameter neural translation model in GGUF quantized format (Q4_K_M). ClariRec runs it locally using the llama.cpp inference engine.
Key facts:
- Size: approximately 1.1 GB on disk
- Download: one-time, from Hugging Face; supports pause and resume
- Verification: SHA-256 checksum after download
- Privacy: after download, translated text never leaves your Mac
- Cost: free — the model, the download, and unlimited local usage
- No account: no Hugging Face account, no ClariRec account, no API key
The model supports all 11 target languages ClariRec offers: Simplified Chinese, Traditional Chinese, English, Japanese, Korean, French, German, Spanish, Russian, Portuguese, and Arabic.
How to set up
- Open ClariRec from the menu bar → Settings → Translation.
- Find the Local Translation Model section.
- Click Download. The progress bar shows download state.
- If you need to pause, click Pause. The download resumes from where it stopped — even if you close Settings or restart ClariRec.
- After download completes, ClariRec verifies the file with a SHA-256 checksum.
- Once verified, the model status shows Ready.
The model is stored in ~/Library/Application Support/ClariRec/Models/. You can delete it from Settings at any time to reclaim disk space.
Disk space: ClariRec checks that at least 1.5× the model size is available before starting the download.
How to use
Once the model is downloaded and ready:
- Set Local Translation as your preferred translation engine in Settings → Translation.
- Translate as you normally would:
- In the result window after OCR (manual or Auto Translate)
- Via macOS Services (right-click selected text → Services → Translate with ClariRec)
- Via the clipboard hotkey (
⌥Tby default)
- All translation happens on your Mac. No network request is made.
Local Translation in the OCR result window: English source above, Simplified Chinese below. The engine label confirms the on-device model — not a BYOK cloud provider.
The local model works with all three translation entry points and follows the same workflow as any other engine — the only difference is that content stays on your device.
How it runs on your Mac
ClariRec loads the model into memory on the first translation request. On Apple Silicon Macs, inference runs on the GPU via Metal for best performance. If GPU loading fails, it falls back to CPU.
To keep things fast:
- Prewarm: when the local model is your preferred engine, ClariRec begins loading it in the background while OCR is still running, so translation is ready the moment OCR finishes.
- Stay resident: the model stays in memory between translations to avoid reload delays. After 5 minutes of inactivity, it unloads automatically.
- Memory pressure: if macOS reports memory pressure, the model unloads immediately to free RAM.
Recommended minimum memory: 4 GB. Most modern Macs (Apple Silicon or recent Intel) handle this comfortably.
Streaming output
The local model supports streaming: translated text appears word by word in the result window or Mini HUD as it is generated, rather than waiting for the entire translation to complete. This gives immediate visual feedback, especially for longer text.
When the local model is a good fit
Choose the local model when:
- You want translation that works without any network connection
- You translate frequently and do not want per-request API costs
- Privacy is a priority — text stays on your Mac
- You need a simple setup with no API key to manage
- You translate between the 11 supported languages regularly
When another engine may be better
The local model is general-purpose. For some tasks, a different engine performs better:
| Scenario | Consider |
|---|---|
| Specialized or domain-heavy text where quality matters most | A BYOK cloud provider (DeepL, OpenAI, Anthropic, etc.) |
| macOS 26+ system with Apple Intelligence ready | Apple Intelligence — also local, no download needed |
| Quick translation on macOS 15+ without any setup | macOS system translation (may download a language pack) |
| A specific language pair where a cloud engine has known strengths | BYOK with that provider |
You can switch engines at any time in the result window or in Settings. Switching is instant — ClariRec retranslates the current text with the new engine.
Technical text protection
When translating text that contains URLs, file paths, inline code, email addresses, or identifiers, ClariRec’s technical text protection (enabled by default) masks these fragments before sending them to the model and restores them afterward. This prevents the model from mangling technical strings during translation.
If a protected fragment is lost in translation output, ClariRec retries the translation without masking as a fallback.
This protection applies to the local model, Apple Intelligence, and all Chat Completions-based cloud providers. DeepL and system translation handle formatting through their own mechanisms.
Compared with other local options
| Option | Setup | Network after setup | Runs in ClariRec |
|---|---|---|---|
| ClariRec local model | One click download | No | Yes — integrated with OCR, Services, ⌥T |
| Apple Intelligence | macOS 26+, supported device, system enabled | No | Yes |
| macOS system translation | macOS 15+, language pack download | Usually no after pack install | Yes (view-layer integration) |
| Standalone local translator | Separate app | Varies | No — separate workflow |
The ClariRec local model is the only option that is available on macOS 14+, needs no system-level feature (no Apple Intelligence, no macOS 15), and integrates directly into the OCR → translate workflow.
Privacy boundary
| What | Where |
|---|---|
| Model file | ~/Library/Application Support/ClariRec/Models/ |
| Model download | One-time from Hugging Face (HTTPS) |
| Translation content | Stays on your Mac — never sent to any server |
| API keys | Not required |
| Analytics | None — ClariRec has no analytics SDKs |
After the initial download, the local model is fully self-contained. ClariRec does not phone home, does not send usage data, and does not proxy your text through any server.
Full privacy boundary: Local OCR for macOS · Privacy Policy
Troubleshooting
Download stalls or fails
Check your network connection. The download supports pause and resume — you can pause, quit ClariRec, and resume later. If the file appears corrupt, delete it from Settings and redownload.
Translation is slow
The first translation after a cold start loads the model into memory and warms up the Metal pipeline. Subsequent translations are faster. On older Macs with limited memory, the model may fall back to CPU inference.
Output is truncated
Very long texts may exceed the model’s context window. Split long documents into shorter sections and translate each one.
You want to free disk space
Delete the model from Settings → Translation → Local Translation Model → Delete. You can redownload it anytime.
Related
Translate screenshot text · Translate selected text · Offline OCR and translation · Local OCR · FAQ
