license: apache-2.0
base_model:
- google/gemma-4-E2B-it
tags:
- litert-lm
litert-community/gemma-4-E2B-it-litert-lm
Main Model Card: google/gemma-4-E2B-it
This model card provides the Gemma 4 E2B model in a way that is ready for deployment on Android, iOS, Desktop, IoT and Web.
Gemma is a family of lightweight, state-of-the-art open models from Google, built from the same research and technology used to create the Gemini models. This particular Gemma 4 model is small so it is ideal for on-device use cases. By running this model on device, users can have private access to Generative AI technology without even requiring an internet connection.
These models are provided in the .litertlm format for use with the LiteRT-LM framework. LiteRT-LM is a specialized orchestration layer built directly on top of LiteRT, Google’s high-performance multi-platform runtime trusted by millions of Android and edge developers. LiteRT provides the foundational hardware acceleration via XNNPack for CPU and ML Drift for GPU. LiteRT-LM adds the specialized GenAI libraries and APIs, such as KV-cache management, prompt templating, and function calling. This integrated stack is the same technology powering the Google AI Edge Gallery showcase app.
LiteRT-LM uese the state of the art Gemma-4 mobile quantization scheme that uses a mixture of 2bit, 4bit and 8 bit weights. This means that for text only use cases the weight footprint in memory can be as low as 0.8 GB while the runtime uses memory mapping to support the 1.12GB of embedding parameters. This approach gives significant working memory savings on some platforms as seen in the more detailed data below. Additionally the Vision and Audio models are loaded on demand to further reduce memory consumption.
Try Gemma 4 E2B
Build with Gemma 4 E2B and LiteRT-LM
Ready to integrate this into your product? Get started here.
Gemma 4 E2B Performance on LiteRT-LM
All benchmarks were taken using 1024 prefill tokens and 256 decode tokens with a context length of 2048 tokens via LiteRT-LM. The model can support up to 32k context length. The inference on CPU is accelerated via the LiteRT XNNPACK delegate with 4 threads. Time-to-first-token does not include load time. Benchmarks were run with caches enabled and initialized. During the first run, the latency and memory usage may differ. Model size is the size of the file on disk.
CPU memory was measured using, rusage::ru_maxrss on Android, Linux and Raspberry Pi, task_vm_info::phys_footprint on iOS and MacBook and process_memory_counters::PrivateUsage on Windows.
Android
Note: On supported Android devices, Gemma 4 is available through Android AI Core as Gemini Nano, which is the recommended path for production applications.
| Device | Backend | Prefill (tokens/sec) | Decode (tokens/sec) | Time-to-first-token (sec) | Model size (MB) | CPU Memory (MB) |
|---|---|---|---|---|---|---|
| S26 Ultra | CPU | 557 | 46.9 | 1.8 | 2583 | 1733 |
| S26 Ultra | GPU | 3,808 | 52.1 | 0.3 | 2583 | 676 |
🚨 NEW: Android with Speculative Decoding 🚨
The numbers in this section include speculative decoding. Speculative decoding is an optimization that accelerates LLMs by using a small, fast "draft" model to quickly predict multiple upcoming tokens, while a larger “target” model then verifies those tokens in parallel. The effectiveness of speculative decoding is task dependent because the “draft” model can more easily predict the correct tokens of some tasks. The metrics in this section were collected from a variety of sample prompts and grouped into categories by task type. The baseline measurements are an average across all task types. The number of input and output tokens varied across prompts. Note that if you download this model before May 5, 2026, you should re-download the model if you want to use speculative decoding. Speculative decoding is available on CPU and GPU on Mobile and Desktop.
| Device | Backend | Task Type | Speculative Decoding? | Decode (tokens/sec) | CPU Memory (MB) |
|---|---|---|---|---|---|
| S26 Ultra | CPU | Baseline | No | 40.7 | 1362 |
| S26 Ultra | CPU | Summarize text | Yes | 47.5 | 1582 |
| S26 Ultra | CPU | Code snippet | Yes | 36.3 | 1440 |
| S26 Ultra | CPU | Rewrite tone | Yes | 47.1 | 1408 |
| S26 Ultra | CPU | Free form | Yes | 38.1 | 1459 |
| S26 Ultra | GPU | Baseline | No | 51.5 | 791 |
| S26 Ultra | GPU | Summarize text | Yes | 91.7 | 817 |
| S26 Ultra | GPU | Code snippet | Yes | 84.4 | 788 |
| S26 Ultra | GPU | Rewrite tone | Yes | 87.4 | 762 |
| S26 Ultra | GPU | Free form | Yes | 66.5 | 804 |
iOS
| Device | Backend | Prefill (tokens/sec) | Decode (tokens/sec) | Time-to-first-token (sec) | Model size (MB) | CPU/GPU Memory (MB) |
|---|---|---|---|---|---|---|
| iPhone 17 Pro | CPU | 532 | 25.0 | 1.9 | 2583 | 607 |
| iPhone 17 Pro | GPU | 2,878 | 56.5 | 0.3 | 2583 | 1450 |
Linux
| Device | Backend | Prefill (tokens/sec) | Decode (tokens/sec) | Time-to-first-token (sec) | Model size (MB) | CPU Memory (MB) |
|---|---|---|---|---|---|---|
| Arm 2.3 & 2.8GHz | CPU | 260 | 35.0 | 4.0 | 2583 | 1628 |
| NVIDIA GeForce RTX 4090 | GPU | 11,234 | 143.4 | 0.1 | 2583 | 913 |
macOS
| Device | Backend | Prefill (tokens/sec) | Decode (tokens/sec) | Time-to-first-token (sec) | Model size (MB) | CPU/GPU Memory (MB) |
|---|---|---|---|---|---|---|
| MacBook Pro M4 Max | CPU | 901 | 41.6 | 1.1 | 2583 | 736 |
| MacBook Pro M4 Max | GPU | 7,835 | 160.2 | 0.1 | 2583 | 1623 |
Windows
| Device | Backend | Prefill (tokens/sec) | Decode (tokens/sec) | Time-to-first-token (sec) | Model size (MB) | CPU Memory (MB) |
|---|---|---|---|---|---|---|
| Intel LunarLake | CPU | 435 | 29.8 | 2.39 | 2583 | 3505 |
| Intel LunarLake | GPU | 3,751 | 48.4 | 0.29 | 2583 | 3540 |
Web
| Device | Backend | Prefill (tokens/sec) | Decode (tokens/sec) | Time-to-first-token (sec) | Model size (MB) | GPU Memory (MB) |
|---|---|---|---|---|---|---|
| Macbook Pro M4 Max | WebGPU | 4,853 | 73 | 1.09 | 2008 | ~1800 |
- Web on LiteRT-LM uses a specially optimized model for Web because of its unique memory constraints. Currently the model is text-only.
IoT
| Device | Backend | Prefill (tokens/sec) | Decode (tokens/sec) | Time-to-first-token (sec) | Model size (MB) | CPU Memory (MB) |
|---|---|---|---|---|---|---|
| Raspberry Pi 5 16GB | CPU | 133 | 7.6 | 7.8 | 2583 | 1546 |
| Jetson Orin Nano | CPU | 109 | 12.2 | 9.4 | 2583 | 3681 |
| Jetson Orin Nano | GPU | 1,142 | 24.2 | 0.9 | 2583 | 2739 |
| Qualcomm Dragonwing IQ8 (IQ-8275) | NPU | 3,747 | 31.7 | 0.3 | 2967 | 1869 |
- NPU model is benchmarked with 4096 context length
Running Gemma 4 E2B on Web with MediaPipe
You can also run Gemma through MediaPipe LLM Inference Engine. However, this route is currently in maintenance mode. To add it to your existing MediaPipe flow, download the gemma-4-E2B-it-web.task model file and run with our sample web page, or follow the guide to add it to your own app.