LiteRT-LM
An on-device large language model runtime built by Google on top of LiteRT, responsible for model loading, KV cache management, sampling-based decoding, and orchestration of streaming output.
Timeline (last 90 days)
LiteRT-LM's sendMessageAsync(text) returns Flow<Message> for streaming receipt of replies.
The LiteRT-LM documentation notes that engine.initialize() can take a long time (up to 10 seconds) and therefore must be run on a background thread or coroutine.
In LiteRT-LM, the Engine should be loaded once and retained, and a new Conversation created for each chat session.
Google MediaPipe's LLM Inference API (Android, iOS and Web) has entered maintenance-only mode, and new features and optimizations will be concentrated in LiteRT-LM.
LiteRT-LM is Google's language model runtime built on LiteRT, responsible for model loading, KV cache management, sampling decoding and streaming output orchestration.
As of October 2026, the current version of LiteRT-LM on Google Maven is 0.18.0 and has not yet reached 1.0.
.litertlm is a custom container format of LiteRT-LM that packages weights, tokenizer, model metadata and chat template into a single file.
LiteRT-LM 是 Google 面向端侧语言模型推理的运行时，LiteRT 前身为 TensorFlow Lite，Google 于 2024 年将其更名并扩展适用范围
Skim Recap 通过 LiteRT-LM 和 WebGPU 在浏览器端执行模型推理，实现完全本地运行，无需账户、无需托管 LLM API、数据不出本地
All Facts (9)
LiteRT-LM's sendMessageAsync(text) returns Flow<Message> for streaming receipt of replies.
50%UnverifiedThe LiteRT-LM documentation notes that engine.initialize() can take a long time (up to 10 seconds) and therefore must be run on a background thread or coroutine.
50%UnverifiedIn LiteRT-LM, the Engine should be loaded once and retained, and a new Conversation created for each chat session.
50%UnverifiedGoogle MediaPipe's LLM Inference API (Android, iOS and Web) has entered maintenance-only mode, and new features and optimizations will be concentrated in LiteRT-LM.
50%UnverifiedLiteRT-LM is Google's language model runtime built on LiteRT, responsible for model loading, KV cache management, sampling decoding and streaming output orchestration.
50%UnverifiedAs of October 2026, the current version of LiteRT-LM on Google Maven is 0.18.0 and has not yet reached 1.0.
50%Unverified.litertlm is a custom container format of LiteRT-LM that packages weights, tokenizer, model metadata and chat template into a single file.
50%UnverifiedLiteRT-LM 是 Google 面向端侧语言模型推理的运行时，LiteRT 前身为 TensorFlow Lite，Google 于 2024 年将其更名并扩展适用范围
50%UnverifiedSkim Recap 通过 LiteRT-LM 和 WebGPU 在浏览器端执行模型推理，实现完全本地运行，无需账户、无需托管 LLM API、数据不出本地
50%