README.md

metadata

title: Translation Engine API
emoji: 🌍
colorFrom: indigo
colorTo: green
sdk: docker
app_port: 7860
pinned: false
short_description: Multi-model Translation API powered by CTranslate2

🌍 Translation Engine API

A high-performance, multi-engine Translation API built with FastAPI and Scalar.

This Space serves local CPU-optimized machine learning models alongside cloud API providers. It is designed to act as the primary translation backend for an upcoming Django web application, but functions completely as a standalone REST API.

---

Interactive API Documentation

Interactive Scalar documentation is hosted directly on the root endpoint: 👉 Open API Docs & Sandbox

---

Available Translation Engines

1. Local CPU Models (Offline & Private)

Local models are automatically downloaded, quantized to INT8 / GGUF format, and cached in ephemeral storage for fast CPU execution with minimal RAM footprint.

Model ID	Provider Name	Base Weights	Precision	Target Languages / Features
nllb	NLLB-200 (600M)	Meta nllb-200-distilled-600M	int8	Fast general translation (200+ languages)
nllb_1_3b	NLLB-200 (1.3B)	Meta nllb-200-distilled-1.3B	int8	Higher accuracy general translation
afrinllb	AfriNLLB (600M)	AfriNLP/AfriNLLB-12enc-12dec-full-ft-kd	int8	Knowledge-distilled for 15+ African languages
madlad	MADLAD-400 (3B)	google/madlad400-3b-mt	int8	Multilingual translation across 400+ languages
swahili_gemma	Swahili Gemma (1B)	CraneAILabs/swahili-gemma-1b-GGUF	Q4_K_M	Fine-tuned for English ↔ Swahili translation

2. Cloud API Engines (Bring-Your-Own-Key)

Cloud API providers allow accessing large language models. You can either configure environment variables on the server or pass your API key dynamically in the JSON payload per request.

Provider ID	Provider Name	Default Model	Required Client Key
gemini	Google Gemini	gemini-1.5-flash	api_key (Gemini API Key)
groq	Groq LLaMA	llama-3.3-70b-versatile	api_key (Groq API Key)
groq_qwen	Groq Qwen	qwen-2.5-coder-32b	api_key (Groq API Key)

---

How to Use the API

1. Check Available Engines

GET /

2. Translate via Local Model (e.g., AfriNLLB)

POST /translate
Content-Type: application/json

{
    "text": "Is it not a beautiful Tuesday morning?",
    "source": "eng_Latn",
    "target": "swh_Latn",
    "engine": "afrinllb"
  }

3. Translate via Cloud API (Passing API Key dynamically)

If no server-side API key is set, clients can supply their own key directly:

POST /translate
Content-Type: application/json

{
  "provider": "gemini",
  "text": "Good morning, how can I help you today?",
  "src": "en",
  "tgt": "sw",
  "api_key": "YOUR_GEMINI_API_KEY_HERE"
}

---

Environment Variables Configuration

If you are hosting this Space yourself or running it locally, you can set optional environment keys in Space Settings:

# Optional Cloud Keys (If set, clients do not need to provide their own keys)
GEMINI_API_KEY=your_gemini_key
GROQ_API_KEY=your_groq_key

# CTranslate2 / CPU Threading Controls
OMP_NUM_THREADS=8
CT2_COMPUTE_TYPE=int8

---

📜 Citations & Credits

This project builds upon open-source research and model checkpoints from the AI community:

AfriNLLB

@inproceedings{moslem-etal-2026-afrinllb,
    title = "{A}fri{NLLB}: Efficient Translation Models for African Languages",
    author = "Moslem, Yasmin and Wassie, Aman Kassahun and Gizachew, Amanuel",
    booktitle = "Proceedings of the Seventh Workshop on African Natural Language Processing (AfricaNLP)",
    year = "2026",
    publisher = "Association for Computational Linguistics"
}

Swahili Gemma

• Fine-tuned and quantized GGUF weights provided by Crane AI Labs.

Meta NLLB-200 & Google MADLAD-400

• Meta AI: No Language Left Behind: Scaling Human-Centered Machine Translation
• Google Research: MADLAD-400: A Multilingual Machine Translation Dataset and Model