Production-Ready On-Device Inference • Zero Remote Telemetry

Autonomous AI Translation
Directly on Mobile Silicon

Parlex is engineered for privacy-first, zero-latency cross-lingual intelligence. It runs 100% on-device neural models without network requests, subscriptions, or corporate data collection.

Android 8.0 — 16 (API 26-36)
Target: arm64-v8a
Engine: llama.cpp GGUF
Google ML Kit Fast NMT
Qualcomm Adreno OpenCL
License: MIT

Hybrid Dual-Tier Translation Controller

Balanced between microsecond UI responsiveness and deep neural linguistic precision.

Tier 1 • Instant Engine ~30ms Latency

Fast NMT On-Device

Google ML Kit distilled neural translation engine for everyday interactions, instant clipboard monitoring, and real-time screen overlay translation.

  • 59 Languages covering Eurasia, Americas, and Middle East
  • 0 MB LLM RAM — ultra battery efficient, zero background footprint
  • .parlex-fast single-archive offline package distribution
Tier 2 • Deep Intelligence Tencent Hy-MT

Primary Neural LLM (JNI)

Tencent's translation-specialized models running via llama.cpp C++17 JNI. Trained specifically for nuance, cross-lingual context, and idiom preservation.

  • 33 Languages & 1056 Directions with high translation fidelity
  • Qualcomm Adreno OpenCL GPU neural tensor acceleration
  • Q4_K_M Quantization (1.13 GB footprint / ~1.8 GB RAM)

Supported Neural Models & Runtime Specs

Transparent memory consumption, quantization profiles, and license requirements.

Model Name Architecture Parameters Quant Download Size Target RAM Runtime Engine
Hy-MT 1.5 Encoder-Decoder LLM 1.8B Q4_K_M 1.13 GB ~1.8 GB llama.cpp JNI (C++17)
Hy-MT2 Mobile Specialized Translation 1.8B Q4_K_M 1.13 GB ~1.8 GB llama.cpp / Adreno OpenCL
Google ML Kit NMT Distilled Compact NMT Edge NMT INT8 35 MB 0 MB Google ML Kit Native
TranslateGemma 4B Multimodal Edge 4.0B E2B / E4B 2.40 GB ~2.8 GB Google LiteRT-LM GPU
PP-OCRv6 Tiny Text Detection & OCR Mobile Vision FP16 18 MB ~80 MB Alibaba MNN Engine

Perception & Linguistic Suite

Built for real-world travel, speech dialogue, and script transliteration without network access.

Speech & Acoustics

Voice Dialogue with Silero VAD

Natural two-way voice conversations with real-time speech turn detection, acoustic echo cancellation (AEC), and Whisper Tiny neural acoustic recognition.

  • Silero Voice Activity Detection (VAD) with instant turn cutoff
  • Hardware AEC guard preventing loopback audio echo
  • Background audio recording (AAC 48 kbps / 16-bit PCM WAV)
Linguistics

ICU 74 Transliteration Engine

Comprehensive offline phonetic romanization engine. Instantly converts non-Latin scripts into readable phonetic transcriptions.

  • Cyrillic to Latin romanization (GOST / ISO standards)
  • CJK support: Chinese Pinyin and Japanese Romaji
  • Arabic, Devanagari (Hindi), Thai, and Korean Hangul scripts

33 Languages Supported Out-of-the-Box

Zero internet required. Full bidirectional coverage across 1056 language pairs.

🇷🇺 Russian
🇺🇸 English
🇨🇳 Chinese
🇪🇸 Spanish
🇩🇪 German
🇫🇷 French
🇯🇵 Japanese
🇰🇷 Korean
🇮🇹 Italian
🇵🇹 Portuguese
🇹🇷 Turkish
🇻🇳 Vietnamese
🇦🇪 Arabic
🇮🇳 Hindi
🇹🇭 Thai
🇮🇩 Indonesian
🇵🇱 Polish
🇳🇱 Dutch
🇬🇷 Greek
🇨🇿 Czech
🇭🇺 Hungarian
🇷🇴 Romanian
🇸🇪 Swedish
🇩🇰 Danish
🇫🇮 Finnish
🇳🇴 Norwegian
🇺🇦 Ukrainian
🇮🇱 Hebrew
🇲🇾 Malay
🇵🇭 Tagalog
🇮🇷 Persian
🇧🇩 Bengali
🇬🇪 Georgian

Model Catalog & Sync API

Automated manifest endpoints for package validation, download verification, and checksums.

Automated Manifest Endpoint

Parlex client applications query the lightweight catalog manifest to verify model weights, SHA-256 hashes, and package updates:

GET /v1/models/catalog.json HTTP/1.1 Host: api.107-173-25-104.sslip.io Accept: application/json
View Live JSON Manifest →