ArchestackGitHub ↗ (opens in a new tab)

Open-source alternatives to DeepL

pdfmathtranslate leads with 64, ahead of libretranslate (58) and rtranslator (45). Updated from public GitHub data.

The highest-scoring open-source projects in Model serving. They are picks, not exact replacements, so read the weaknesses before you switch. The same picks fit if you are leaving DeepL Translator or Google Translate.

Permissive license
rtranslator Apache-2.0
Open-source alternatives to DeepL
RankProjectScore
1pdfmathtranslateTranslates scientific PDFs while keeping formulas, figures and layout intact Live demo ↗ (opens in a new tab)Data and scraping for AI37.7k starsAGPL-3.064 out of 100
2libretranslateSelf-hosted machine translation API built on Argos Translate Live demo ↗ (opens in a new tab)Model serving17.1k starsAGPL-3.058 out of 100
3rtranslatorOffline real-time speech translator for Android using on-device Whisper and NLLBVoice10.5k starsApache-2.045 out of 100

Reviews

164 out of 100

pdfmathtranslate

Translates scientific PDFs while keeping formulas, figures and layout intact

Live demo ↗ (opens in a new tab)37.7k starsAGPL-3.0last commit Oct 2026

pdf2zh translates scientific PDFs and writes a translated-only and a bilingual (dual) PDF, preserving formulas, charts, table of contents and annotations. It runs as a CLI, a Gradio web UI on port 7860, a Python or HTTP API, or an MCP server, and calls a configurable translation service (Google by default). Layout detection uses a DocLayout-YOLO ONNX model, with optional OCR and an --ultrafast mode.

Strengths

  • Outputs both translated-only and side-by-side bilingual PDFs from one command
  • Many interfaces: CLI, Gradio GUI, Docker, Python API, HTTP API, MCP stdio/SSE
  • Selectable translation backends via -s, with Google as the default
  • Optional OCR for scanned pages and --ultrafast mode for native-text PDFs

Weaknesses

  • Python limited to 3.11 to 3.12
  • OCR is experimental and targets white-background scans; handwriting may fail
  • Last release is 2025-07-11; v2 work lives in a separate PDFMathTranslate-next repo
  • --ultrafast is CLI fast mode only and is weaker on complex tables or formulas
  • no GPU
  • Docker + Compose
  • Needs DocLayout-YOLO ONNX model (Hugging Face download), Translation service (Google default, others configurable)
  • Models: ONNX
  • port 7860
Live demo ↗ (opens in a new tab)Documentation ↗ (opens in a new tab)Repository on GitHub ↗ (opens in a new tab)
258 out of 100

libretranslate

Self-hosted machine translation API built on Argos Translate

Live demo ↗ (opens in a new tab)17.1k starsAGPL-3.0last commit Oct 2026

LibreTranslate is a self-hosted machine translation API with a web interface for translating text. Translation runs locally on the open-source Argos Translate engine, so it does not call Google, Azure or other proprietary services. The README is short and defers setup and API usage to the external docs site.

Strengths

  • Translates locally through Argos Translate, with no proprietary translation API calls
  • Public hosted instance at libretranslate.com to try before installing
  • Separate API usage guide and community forum
  • Written in Python and published on PyPI

Weaknesses

  • README gives no hardware requirements, ports or language list
  • Setup and API details live only on the external docs site
  • AGPL-3.0 license imposes obligations on networked modifications
  • Translation quality depends on Argos models, which the README does not benchmark
  • Docker + Compose
  • Needs Argos Translate
  • Models: Argos Translate
Live demo ↗ (opens in a new tab)Documentation ↗ (opens in a new tab)Website ↗ (opens in a new tab)Repository on GitHub ↗ (opens in a new tab)
345 out of 100

rtranslator

Offline real-time speech translator for Android using on-device Whisper and NLLB

10.5k starsApache-2.0last commit Oct 2026

RTranslator is an Android app that translates speech and text on the phone using Whisper for speech recognition and NLLB for translation, both converted to ONNX and int8-quantized. Conversation mode links phones over Bluetooth LE, with Bluetooth headset support, and WalkieTalkie mode handles two languages on a single phone. Text-to-speech uses the system TTS engine. The 3.0 beta replaces NLLB with Bergamot, Madlad 400 3B or HY-MT 1.5 1.8B.

Strengths

  • Runs speech recognition and translation on the phone, usable offline with no server
  • Conversation mode supports more than two phones and Bluetooth headsets
  • WalkieTalkie mode works with one phone and detects which of two languages is spoken
  • Models are ONNX int8 with KV cache; NLLB uses 1.3 GB RAM versus 2.5 GB baseline

Weaknesses

  • Needs at least 6 GB RAM and a fast CPU; weaker phones may crash
  • First launch downloads 1.2 GB of models from GitHub, slow in some regions
  • Android only; README lists Bluetooth connection drops as a known bug
  • NLLB is non-commercial licensed and Google ML Kit is closed source in the 2.x README
  • RAM ≥ 6 GB
  • no GPU
  • Needs Android system TTS (Google TTS recommended), ONNX Runtime, Bluetooth LE
  • Models: Whisper-Small-244M (ONNX int8), NLLB-Distilled-600M (ONNX int8), Mozilla Bergamot (3.0 beta), Madlad 400 3B (3.0 beta), HY-MT 1.5 1.8B (3.0 beta)
Documentation ↗ (opens in a new tab)Repository on GitHub ↗ (opens in a new tab)