Pantogloss

Pantogloss is a TensorFlow/Keras library that translates text from many languages into English. It provides a Python API, command-line interface, and optional persistent local server with a browser UI.

The default public model is pantogloss-500-en-compact, a 538 MiB model validated on NVIDIA CUDA and Apple Metal. Model weights are downloaded separately from Hugging Face and are never bundled in the Python wheel. Public models normally require no Hugging Face login.

What it can do

• Translate individual strings, batches, files, and bounded input streams.
• Run on CPU, NVIDIA CUDA, or Apple Silicon Metal.
• Keep a model resident behind a local HTTP API for fast repeated requests.
• Provide a simple Any language → English browser translator.
• Offer fast greedy and higher-quality beam-search decoding presets.
• Return plain strings or structured results with timing and runtime metadata.
• Attach caller-supplied language quality guidance and conservative warnings.
• Preserve input order and isolate individual failures in long-running streams.

Pantogloss translates text; it does not detect the source language or parse document formats. See Project scope.

Installation

Pantogloss supports Python 3.10–3.12. Use the extra for your accelerator:

# CPU
python -m pip install pantogloss

# NVIDIA GPU on Linux
python -m pip install "pantogloss[cuda]"

# Apple Silicon Metal
python -m pip install "pantogloss[metal]"

# Persistent API and browser UI; combine with cuda or metal when needed
python -m pip install "pantogloss[server,metal]"

The first use downloads the selected model. Later loads use the local Hugging Face cache.

Python API

Load one translator and reuse it:

from pantogloss import Translator

translator = Translator.from_pretrained(device="auto")
print(translator.translate("Comment allez-vous ?"))

Translate a batch or request structured results:

translations = translator.translate([
    "Hola señor",
    "Wie geht es Ihnen?",
])

result = translator.translate_detailed("Bonjour le monde.")
print(result.text)
print(result.elapsed_seconds, result.execution_device)

For a large iterable, translate_iter() and translate_iter_detailed() process bounded batches without retaining the complete input or output collection.

with open("source.txt", encoding="utf-8") as source:
    for translation in translator.translate_iter(source, batch_size=16):
        print(translation)

Use the quality preset when latency is less important than beam-search quality:

print(translator.translate("Comment allez-vous ?", preset="quality"))

Command line

pantogloss translate "Comment allez-vous ?"
pantogloss translate --preset quality "Comment allez-vous ?"
pantogloss translate --input source.txt --output english.txt --batch-size 16
pantogloss models
pantogloss info --device gpu --json
pantogloss doctor

Translation output stays clean on stdout. Use --verbose, --report, or --tensorflow-logs when diagnostics are wanted. Use --offline to require an already cached model.

Persistent server and browser UI

Each standalone CLI invocation reloads the model. For repeated requests, keep one warmed translator in memory:

python -m pip install "pantogloss[server,metal]"  # or server,cuda
pantogloss serve --device gpu

Open http://127.0.0.1:8765/ for the two-pane browser translator. The same process exposes:

• GET /health
• GET /info
• GET /metrics
• POST /translate

curl http://127.0.0.1:8765/translate \
  -H 'Content-Type: application/json' \
  -d '{"text":"Comment allez-vous ?"}'

The server binds to loopback, warms the model before reporting readiness, and serializes inference by default. Queuing is bounded, shutdown drains active TensorFlow work, and optional request logs contain timing/count metadata rather than source or translated text. On Apple Silicon, the server defaults to the bounded-memory eager decoder because TensorFlow Metal retains memory during compiled decoding; --metal-compiled-decode is an explicit faster but memory-growing opt-in. See the Server and Web UI wiki page for deployment, authentication, limits, API schemas, benchmarks, and the pantogloss balance front end for independently resident workers. The balancer supports explicit graceful draining, connection-only safe retries, passive circuit breaking, and Prometheus-compatible reliability metrics.

Models

Select a model explicitly when needed:

translator = Translator.from_pretrained("pantogloss-500-en-fp16", device="gpu")

The compact model retains the best tested balance of size, quality, memory, and speed. The INT8 alternative is not the default because it is substantially slower. Detailed evidence and backend caveats live in the wiki and checked-in experiment reports.

Documentation

• Installation and Quickstart
• Server and Web UI
• Decoding Presets
• Platform Validation
• Language Quality Catalog
• Quality Evaluation
• Development and Testing
• Tika Document Translation

The repository also retains reproducible model-conversion, parity, compression, and evaluation artifacts under docs and evaluation.

Project scope

Pantogloss translates text and ordered text segments. It intentionally does not identify languages, detect file types, or parse documents. DocumentTranslator is a neutral text-segmentation helper. pantogloss-tika remains a compatibility example and is not a direction for new core dependencies. See the architecture boundary.

Translation quality varies by language, domain, and input. Pantogloss does not provide calibrated confidence and should not be relied on without review for medical, legal, safety-critical, or other high-stakes decisions.

Provenance and license

The original model was described by Thamme Gowda, Zhao Zhang, Chris A. Mattmann, and Jonathan May in Many-to-English Machine Translation Tools, Data, and Pretrained Models, ACL-IJCNLP 2021 System Demonstrations, DOI 10.18653/v1/2021.acl-demo.37.

Pantogloss and its converted models are licensed under the Apache License, Version 2.0. See NOTICE for attribution.