Welcome to the Google Tensor SDK release notes! Here you'll find information on the latest releases, updates, and improvements for on-device AI development on Google Tensor chipsets. ## SDK v2.1 This release introduces support for a new chipset and models to enhance on-device performance and capabilities. *Released: Oct 14, 2026* ### What's new - ***Embedding Gemma v2*** is now available as a pre-compiled model on both Pixel 10 and Pixel 11. See [Pre-compiled SLMs](https://developers.google.com/edge/tensor-sdk/slm-compilation#precompiled) for more details. - You can now export your fine-tuned checkpoints for Gemma 3 1B, FunctionGemma and Embedding Gemma models directly into `.litertlm` bundles. This lets you run custom-trained models on Tensor G6 TPUs without relying solely on pre-compiled builds. See [Compile SLMs from fine-tuned checkpoints](https://developers.google.com/edge/tensor-sdk/slm-compilation#finetuned) to learn more. ### Improvements - We've added new models to the [Model Garden](https://developers.google.com/edge/tensor-sdk/model-garden) ranging across multiple types and use cases (like depth estimation, face reconstruction, image \& text understanding, and image classification). ## Compatibility See the compatibility information for the Tensor SDK, detailing supported SoCs, LiteRT versions, and Android versions across different SDK releases. | SDK Release | Target SoC | LiteRT version || Android version | | SDK Release | Target SoC | Stable (Onwards) | Nightly | Android version | |---|---|---|---|---| | **v2.1** | Tensor G6 | [`ai-edge-litert 2.3.0`](https://github.com/google-ai-edge/LiteRT/releases#release-v2.3.0) *Released: October 2026* | [`ai-edge-litert-nightly 2.3.0.dev20261001`](https://pypi.org/project/ai-edge-litert-nightly/2.3.0.dev20261001/) *Released: October 2026* | 26D1 | | **v2.1** | Tensor G5 | [`ai-edge-litert 2.3.0`](https://github.com/google-ai-edge/LiteRT/releases#release-v2.3.0) *Released: October 2026* | [`ai-edge-litert-nightly 2.3.0.dev20261001`](https://pypi.org/project/ai-edge-litert-nightly/2.3.0.dev20261001/) *Released: October 2026* | 26Q3 | | **v2.0** | Tensor G6 | [`ai-edge-litert 2.1.6`](https://github.com/google-ai-edge/LiteRT/releases#release-v2.1.6) *Released: July 2026* | [`ai-edge-litert-nightly 2.2.0.dev20260809`](https://pypi.org/project/ai-edge-litert-nightly/2.2.0.dev20260809/) *Released: August 2026* | 26D1 | | **v2.0** | Tensor G5 | [`ai-edge-litert 2.1.6`](https://github.com/google-ai-edge/LiteRT/releases#release-v2.1.6) *Released: July 2026* | [`ai-edge-litert-nightly 2.2.0.dev20260809`](https://pypi.org/project/ai-edge-litert-nightly/2.2.0.dev20260809/) *Released: August 2026* | 26D1 | | **v1.02** | Tensor G5 | [`ai-edge-litert 2.1.6`](https://github.com/google-ai-edge/LiteRT/releases#release-v2.1.6) *Released: July 2026* | [`ai-edge-litert-nightly 2.2.0.dev20260630`](https://pypi.org/project/ai-edge-litert-nightly/2.2.0.dev20260630/) *Released: June 2026* | 26Q2 | | **v1.01** | Tensor G5 | [`ai-edge-litert 2.1.5`](https://github.com/google-ai-edge/LiteRT/releases#release-v2.1.5) *Released: May 2026* | [`ai-edge-litert-nightly 2.2.0.dev20260530`](https://pypi.org/project/ai-edge-litert-nightly/2.2.0.dev20260530/) *Released: May 2026* | 26Q1 | ## SDK v2.0 This release introduces support for a new chipset and models to enhance on-device performance and capabilities. *Released: Sep 2, 2026* ### Improvements - Tensor SDK now supports Tensor G6 chipset, available on Pixel 11 series. - We've added support for [EmbeddingGemma-300m](https://huggingface.co/google/embeddinggemma-300m), a model that facilitates efficient on-device search and retrieval, in our [Model Garden](https://developers.google.com/edge/tensor-sdk/model-garden). The [Tensor FAQs](https://developers.google.com/edge/tensor-sdk/tensor-faqs) are now updated to include more scenarios about AOT compilation and memory constraints. ### Bug fixes - Resolved an internal compiler error that occurred when compiling models like [`wav2vec2_base_960h_179200`](https://developers.google.com/edge/tensor-sdk/model-garden#speech-recognition) for Google Tensor SoCs using the LiteRT compiler plugin. ## SDK v1.02 This release introduces key updates and enhancements for on-device AI development, including a new on-device LiteRT-LM chatbot, enhancements to the Google Tensor Options API, updated documentation, optimizations, and bug fixes for the Tensor chipsets. *Released: Jul 2, 2026* ### What's new - We've created a new on-device [LiteRT-LM chatbot app](https://github.com/google-ai-edge/litert-samples/tree/main/compiled_model_api/google/sample_app_tpu) that demonstrates running multimodal LLMs (for example, Gemma 4) on Google Pixel TPUs. This application lets you leverage default multimodal queries like sending text, picking images from the Gallery, or recording audio, which you can expand upon for your specific use cases or use to optimize your own applications. - You can now better fine-tune TPU power and performance states for your applications using the new options in the [Google Tensor Options](https://github.com/google-ai-edge/LiteRT/blob/main/litert/cc/options/litert_google_tensor_options.h) API. ### Improvements - The [Tensor FAQs](https://developers.google.com/edge/tensor-sdk/tensor-faqs) are now updated to include more scenarios about AOT compilation and memory constraints. ### Bug fixes - Resolved TPU performance degradation on Tensor G5 devices by providing platform-optimized, ahead-of-time (AOT) compiled models, improving inference times to ensure stable execution. ## 🛠 Resources \& support - Tensor SDK Home Page: [Google Tensor: On-device AI with LiteRT](https://ai.google.dev/edge/tensor-sdk/) - Hugging Face: [LiteRT Hugging Face community](https://huggingface.co/collections/litert-community/google-tensor) - Technical Support: [GitHub community](https://github.com/google-ai-edge/LiteRT/issues) - LiteRT GitHub: [LiteRT GitHub repository](https://github.com/google-ai-edge/LiteRT) - LiteRT Documentation: [LiteRT documentation](https://ai.google.dev/edge/litert)