LiteRT Community

LiteRT is Google's on-device framework for high-performance ML & GenAI deployment on edge platforms. It is the improved successor to TensorFlow Lite. On this community page, you can find ready-to-run LiteRT models for a wide range of ML/AI tasks.

Within this ecosystem, LiteRT-LM specializes in cutting edge GenAI. Recognizing that LLMs now function as complex pipelines of related models rather than single standalone models, LiteRT-LM leverages LiteRT to deliver an optimized solution for running LLMs on-device.

Both LiteRT and LiteRT-LM can run on a variety of devices including Android, iOS, Windows, macOS, Linux, IoT and Web allow easy deployment and scaling across a diverse device landscape.

Trying it Live

Not sure where to start? We recommend first trying our models in our Google AI Edge Gallery app on Android and iOS.

🌟 Community's Picks of the Week

• EmbeddingGemma v2 (Embedding): embeddinggemma-2-740m-litert-lm (EmbeddingGemma 2 740M is the full, natively multimodal variant of the EmbeddingGemma 2 family). embeddinggemma-2-text-vision-440m-litert-lm (EmbeddingGemma 2 Text Vision 440M is a lightweight, multimodal (text and vision) variant). embeddinggemma-2-text-270m-litert-lm (EmbeddingGemma 2 Text 270M is a highly efficient, text-only variant). They are optimized for low-latency deployment on consumer hardware and edge devices via LiteRT. Main Model Card: google/embeddinggemma-2.
• Qwen3.8-27B (Text Generation) Qwen's 27B model of the Qwen3.8 release, as one 15.38 GB LiteRT-LM file for desktop-class machines: text in, text out, with thinking off.
• clef-flash-LiteRT (Text Classification) Cloudflare's decision model for yes/no, choice and rating questions about a text, a JSON value or a picture, returning every option's probability, as LiteRT graphs and a Python host for desktop runtimes.
• WeVisOne-VL-2B (Image-Text-to-Text) Tencent's WeVisOne-VL-2B, a vision-language model that answers detection, pointing, counting and OCR requests with boxes and points: one LiteRT-LM file, and LiteRT graphs with the pixel decoder for cutting objects out.
• Qwen3-ASR-1.7B (Automatic Speech Recognition) Qwen's speech recognition model (2.04 B parameters) as one LiteRT-LM file: audio in, text out.
• phonon-31m-en (Text-to-Speech) Gradium's Phonon 31M, an English text-to-speech model, packed with its tokenizer into one .litertlm file.
• GLiNER2.5-Multi-LiteRT (Token Classification) GLiNER2.5 Multi, which extracts person, organization, location, product and date entities from English and Japanese text, here as LiteRT files.

Community Contributions

Are we missing your favorite model? You can convert and run PyTorch, TensorFlow, or JAX models to the classic TFLite format using the LiteRT conversion and optimization tools. Or for LLMs, you can use the LiteRT Torch Generative API. When your model is ready, join the LiteRT community org and upload the model here for others to try!

Submission Guidelines

• Model Conversion: Your model must be converted using the LiteRT-Torch or Generative API in .tflite or .litertlm formats.
• Location: All contributions must be made in the community contribution folder (e.g., LiteRT-Community/community-contribution/mobile-bert).
• Licensing: If your model has a license, it must be either an Apache 2.0 or MIT open-source license. Models with no license are also accepted.
• Data Provenance and Privacy: If applicable, you must include a summary of the model's training data in your submission. Additionally, you must provide confirmation that all Personally Identifiable Information (PII) has been scrubbed from the data.

Policies & Support

• Content Policy: All uploaded models must fully comply with the Hugging Face content guidelines.
• Community Moderation: The Google AI Edge team will not actively moderate community contributions to this community, except for reactive removal in clear cases of spam or abuse.
• Conversion Requests: In the event that you cannot successfully convert a .tflite or .litertlm model, you may submit a PR requesting conversion support from the LiteRT team.