aoiandroid/streamgemma-micro-litert-lm

StreamGemma-Micro LiteRT-LM (mobile-pair)

This is the mobile-optimized 550M parameter student variant of StreamGemma, compressed via dynamic range INT8 quantization and packaged for the LiteRT-LM execution framework.

It is designed specifically for resource-constrained edge devices (e.g., iPhone 12+, Pixel 6+), requiring a minimal runtime memory footprint of approximately 450 MB.

Model Specifications

• Architecture: StreamGemma-Micro (mobile-pair variant)
• Parameters: ~550M (d_model=512, 10 transformer layers)
• Precision: Dynamic range INT8 quantization
• Vocab Size: 256,000
• Supported Languages: Tier S priority languages (eng, jpn, cmn)
• Key Pipeline Features:
  • Auto-regressive Text Prefill & Generation
  • Language Identification (LID)
  • Speculative Decoding support

Text-Only Translation Model: This Micro variant is designed specifically for text-to-text translation (machine translation) on-device. It does not contain an audio_model or a TTS (text-to-speech) head to keep its size and memory footprint optimal. For end-to-end speech-to-speech, please refer to the unreduced 2.2B models.

Bundle Files

The repository contains:

1. streamgemma-micro.litertlm: The official unified LiteRT-LM model container file packed using litert-lm-builder.
2. README.md: Model card.
3. streamgemma_banner.jpg: Banner asset.

Deployment with streamgemma-litert-lm

This model can be run directly using the cross-platform msandroid/streamgemma-litert-lm SDKs (Python, Android Kotlin, iOS Swift, Web TypeScript, Flutter, C++, Rust).

Python Example

import numpy as np
from streamgemma_litert import StreamGemmaEngine, EngineConfig, SessionConfig

# Initialize the engine with the .litertlm file
config = EngineConfig(model_path="path/to/streamgemma-micro.litertlm")
engine = StreamGemmaEngine(config)

# Start a translation session
session = engine.create_session(SessionConfig())

# Feed input tokens (shape [1, 128])
input_tokens = np.array([[2, 101, 102, 103] + [0]*124], dtype=np.int32)
result = session.feed_text(input_tokens)
print(f"[{result.source_language}] {result.text}")