LibreTranslate is the only serious open-source, self-hosted translation engine. For teams that can't send data to external APIs (due to compliance, privacy, or policy requirements), it's the primary option. But "open-source and self-hosted" doesn't mean "free." There are real costs, real quality trade-offs, and real maintenance burdens.
I tested LibreTranslate against Google Translate and Langbly on the same set of translation tasks. Here's what I found.
What LibreTranslate is
LibreTranslate is an open-source translation API built on the Argos Translate engine. It runs as a Docker container or standalone Python application. You deploy it on your own server and call it via REST API, just like you'd call Google Translate or Langbly.
The project is maintained by a small team of contributors and funded primarily through the hosted version (LibreTranslate.com) and donations. The core engine uses CTranslate2 for inference and trains on the OPUS parallel corpus.
Quick comparison
| Feature | LibreTranslate | Google Translate | Langbly |
|---|---|---|---|
| Deployment | Self-hosted | Cloud API | Cloud API |
| Cost model | Server costs | $20/M chars | $1.99-3.80/M chars |
| Languages | ~30 | 130+ | 130+ |
| Quality tier | Basic NMT | Standard NMT | Context-aware |
| Data privacy | Full control | Google processes | Processed, not stored |
| Maintenance | You | None | None |
| SLA | None | 99.5%+ | 99.5%+ |
| HTML support | Limited | Full | Full |
| Batch translation | Manual batching | Built-in | Built-in |
Translation quality
I tested all three on 100 segments across four content types: software UI strings, documentation paragraphs, marketing copy, and error messages. Languages tested: English to German, Spanish, French, Japanese, and Dutch.
Results summary
For European languages (DE, ES, FR, NL):
- Langbly: Best overall. Handled placeholders and technical terms well. Natural phrasing.
- Google Translate: Good accuracy, occasionally stiff phrasing. Reliable with HTML and placeholders.
- LibreTranslate: Understandable but noticeably weaker. More grammatical errors, awkward word choices, and occasional meaning shifts.
For Japanese:
- Langbly: Good accuracy, appropriate formality level.
- Google Translate: Decent but sometimes chose the wrong formality register.
- LibreTranslate: Struggled significantly. Several mistranslations that changed meaning. Japanese is a low-resource language for the Argos engine.
The quality gap is most visible in complex sentences. Simple phrases ("Save changes," "Loading...") translate acceptably on all three. But a sentence like "Your subscription has been updated and will take effect at the start of your next billing cycle" exposes the differences. LibreTranslate's Argos engine doesn't handle relative clauses and domain-specific vocabulary as well as the larger engines.
Placeholder handling
This is where LibreTranslate struggles most in software localization contexts. Strings like Welcome, {name}! You have {count} unread messages. often come back with mangled placeholders. The Argos engine doesn't have built-in placeholder protection, so {name} might become {nom} or disappear entirely.
Google Translate and Langbly both handle placeholders reliably when you specify the correct format parameter.
True cost of self-hosting
LibreTranslate is free to download and run. But "free software" has costs:
Server costs
LibreTranslate needs a reasonably powerful server. The language models consume memory, and translation requests need CPU (or GPU for faster inference).
- Minimum viable: 4GB RAM, 2 CPU cores. Handles ~5 requests/second. ~$20-30/month on most cloud providers.
- Production: 8-16GB RAM, 4 CPU cores. Handles ~15-20 requests/second. ~$50-100/month.
- High throughput: 32GB RAM, 8 cores, or GPU instance. Handles 50+ requests/second. ~$200-500/month.
Loading all 30+ language models requires significant memory. If you only need 5-6 languages, you can reduce memory by loading only those models.
Maintenance costs
Self-hosting means you handle:
- OS and security updates
- Docker/container management
- Monitoring and alerting
- Scaling during traffic spikes
- Backup and disaster recovery
- LibreTranslate version upgrades (model updates improve quality but require testing)
Budget 2-5 hours per month for a healthy LibreTranslate installation. At a developer's hourly rate, that's $200-1,000/month in time.
Total cost comparison
For a team translating 25 million characters per month:
| Service | Monthly cost | Notes |
|---|---|---|
| LibreTranslate | $50-100 server + $200-1000 time | Lower quality, maintenance burden |
| Google Translate | $490 | No maintenance, reliable quality |
| Langbly Growth | $69 | No maintenance, context-aware quality |
LibreTranslate is only cheaper if you value your time at zero and accept the quality trade-off. For most teams, the maintenance burden makes it more expensive than a managed API.
When LibreTranslate makes sense
Despite the drawbacks, there are legitimate reasons to choose LibreTranslate:
Data sovereignty requirements
Some organizations cannot send text to external APIs under any circumstances. Government agencies, defense contractors, healthcare companies handling patient data, and organizations subject to strict data residency laws. If your data cannot leave your network, LibreTranslate (or a similar self-hosted solution) is your only option.
Note: Langbly offers an EU data residency option that processes translations within the EU. This satisfies many GDPR-related requirements without the overhead of self-hosting.
Offline environments
If you need translation in air-gapped environments (ships, remote facilities, military installations), self-hosted is the only path. LibreTranslate runs without internet access once the models are downloaded.
Research and experimentation
For academic research, testing translation approaches, or building educational tools, LibreTranslate's open-source license (AGPL-3.0) allows free use. The ability to inspect and modify the engine is valuable for research purposes.
Very low volume
If you're translating a few hundred strings per month and already have a server running with spare capacity, adding LibreTranslate as a Docker container costs effectively nothing. The quality trade-off matters less for internal tools or prototypes.
When LibreTranslate doesn't make sense
- Production software with user-facing translations: The quality gap produces translations that users notice. For a professional product, it reflects poorly.
- High volume: The infrastructure and maintenance costs exceed managed API costs.
- Many languages: LibreTranslate supports ~30 languages. If you need Vietnamese, Thai, Filipino, or dozens of other languages, it can't help you.
- Speed-critical applications: Latency depends on your server capacity. A $30/month VPS translates much slower than a dedicated API endpoint backed by optimized infrastructure.
Migration path
If you're currently using LibreTranslate and want to switch to a managed API, the migration is straightforward:
- API compatibility: LibreTranslate uses a simple
POST /translateendpoint. Langbly uses the Google Translate v2 format (POST /language/translate/v2). The request structure is different but both are simple REST calls. Updating the API call in your code takes 15-30 minutes. - Language codes: LibreTranslate uses ISO 639-1 codes (same as Google and Langbly). No mapping needed.
- Testing: Run your existing test strings through the new API and compare output. You'll likely see quality improvements across all language pairs.
If you're evaluating LibreTranslate versus a managed API for a new project, start with the Langbly free tier (500K characters). Test with your actual content. The quality and convenience difference is usually enough to make the decision obvious.