aoiandroid/streamgemma-2.2b-litert-lm

StreamGemma-2.2B LiteRT-LM

This is the quantized, LiteRT-LM format package of the unreduced 2.21B parameter StreamGemma model. The model features dynamic range INT8 quantization and is optimized for ultra-low latency streaming inference on edge devices (Android, etc.).

Model Details

• Architecture: StreamGemma-Nano v2
• Parameters: 2.21B backbone
• Format: LiteRT / TFLite Flatbuffers (INT8 Quantized)
• Component Files:
  • audio_model.tflite: Handles PCM audio inputs (1.6GB)
  • text_model.tflite: Handles tokenized text inputs (2.6GB)
  • vision_model.tflite: Handles vision patch embeddings (2.6GB)
  • draft_model.tflite: Speculative decoding draft module (2.6GB)