universal-translator

A universal text translator built in Rust. Uses llama.cpp (via the llama-cpp-2 Rust crate) for fast, fully local inference of TranslateGemma 4B (Gemma 3 4B instruction-tuned), covering 55 languages, and Lingua for automatic source-language detection.

No API keys required. No network calls at runtime. Everything runs on your machine.

License: MIT OR Apache-2.0 — see LICENSE-MIT and LICENSE-APACHE. Model attributions: see ATTRIBUTIONS.md — covers TranslateGemma 4B and the runtime libraries used for inference. Model license: TranslateGemma 4B weights are subject to the Gemma Terms of Use — see LICENSE-GEMMA and NOTICE.

---

Supported languages

55 supported languages:

Code	Language	Code	Language	Code	Language
af	Afrikaans	hr	Croatian	pt	Portuguese
am	Amharic	hu	Hungarian	ro	Romanian
ar	Arabic	id	Indonesian	ru	Russian
bg	Bulgarian	it	Italian	si	Sinhala
bn	Bengali	ja	Japanese	sk	Slovak
ca	Catalan	kn	Kannada	sl	Slovenian
cs	Czech	ko	Korean	sr	Serbian
da	Danish	lt	Lithuanian	sv	Swedish
de	German	lv	Latvian	sw	Swahili
el	Greek	ml	Malayalam	ta	Tamil
en	English	mr	Marathi	te	Telugu
es	Spanish	ms	Malay	th	Thai
et	Estonian	mt	Maltese	tr	Turkish
fa	Persian	ne	Nepali	uk	Ukrainian
fi	Finnish	nl	Dutch	ur	Urdu
fr	French	no	Norwegian	vi	Vietnamese
gu	Gujarati	pa	Punjabi	yi	Yiddish
ha	Hausa	pl	Polish	zh	Chinese
hi	Hindi

Source language is detected automatically — no configuration required.

---

Workspace layout

universal-translator/
├── translator-core/   # Core library: engine, language detector, types
├── translator-api/    # Axum HTTP API server
├── translator-cli/    # Command-line interface
└── docs/models.md     # Model management guide

Prerequisites

• Rust toolchain (stable) — rustup.rs
• CMake ≥ 3.14 and a C++17 compiler (required to build vendored llama.cpp)
  • macOS: xcode-select --install (provides both)
  • Ubuntu/Debian: sudo apt install cmake g++
  • Fedora: sudo dnf install cmake gcc-c++
• Tested on Linux (x86_64, arm64) and macOS (Apple Silicon)

Quick start

Build

cargo build --release

Get the model

cargo run -p translator-cli -- setup

This downloads TranslateGemma 4B Q8_0 (~4.1 GB) directly from HuggingFace into the default model directory. No Python or HuggingFace CLI required.

For the smaller Q4_K_M variant (~2.6 GB):

cargo run -p translator-cli -- setup --url https://huggingface.co/mradermacher/translategemma-4b-it-GGUF/resolve/main/translategemma-4b-it.Q4_K_M.gguf

Use --model-path <path> to select a specific model file at runtime. Q8_0 is the default. See docs/models.md for details and alternative hosting options.

Run the API server

cargo run -p translator-api

The server listens on http://localhost:3000. See API.md for endpoint reference and request/response schemas.

To run with full observability (traces, metrics, logs pushed to a local monitoring stack):

docker compose -f docker/docker-compose.yml up -d
OTLP_ENDPOINT=http://localhost:4317 \
  cargo run -p translator-api --features opentelemetry

Grafana opens at http://localhost:3001 with a pre-built dashboard. See METRICS.md for the full setup guide and metrics catalogue.

Run the CLI

cargo run -p translator-cli -- translate -t "Hello world" -l fr

See CLI.md for the full command reference (translate, detect-language, detect, languages) including all flags and output formats.

Documentation

• CLI Reference — all subcommands, flags, and output formats
• API Reference — HTTP endpoints, request/response schemas, examples
• Engine Internals — inference limits, concurrency, sampling, language detection
• Observability — OTel traces/metrics/logs, Grafana dashboard, metrics catalogue

Licensing

The source code in this repository is dual-licensed under the Apache License, Version 2.0 and the MIT License. You may choose either license for your use of the source code.

Model license

The TranslateGemma model weights and any model derivatives are not covered by the Apache or MIT licenses. They are subject to the Gemma Terms of Use and Gemma Prohibited Use Policy. By downloading or using the model, you agree to abide by those terms. See LICENSE-GEMMA and NOTICE.