Google launches ultra-small AI model Gemma 3 270M! Cell phones can run it, a new breakthrough for smart devices running offline!

Google launches smallest everGemma 3Open source model, a lightweight model with 270 million parameters, with an underlying design aimed at task-specific fine-tuning, with robust instruction tracking and text capabilities.

In the Instruction Execution Capability test, as shown in the IFEval benchmark, the Gemma 3 270M outperforms the much larger Qwen2.5 0.5B Instruct and matches the Llama 3.2 1B capability.

The Gemma 3 270M is capable of meeting or even exceeding the capabilities of the larger models on a number of specific tasks. The size and performance of the model make it veryIdeal for offline, web-based creative tasks. For example, Google has published a case study of its use of Gemma 3 270M to power a bedtime story generator web app using Transformers.js, which generates great bedtime stories by simply checking a box.

The core capabilities of the Gemma 3 270M are as follows:

1. Lightweight but powerful architecture.The model has a total of 270 million parameters, 170 million embedding parameters and 100 million Transformer module parameters due to the large vocabulary. Thanks to the large vocabulary of 256k tokens, the model is able to handle specific and rare tokens, making it a high-quality base model that can be further fine-tuned in specific domains and languages.

2. Extreme energy efficiency. A major advantage of the model is its low power consumption, and its internal testing on the Pixel 9 Pro SoC showed that the INT4 quantization modelConsumes only 0.751 TP4T in 25 conversations, making it Google's most energy-efficient Gemma model.

3. Instructions are followed.The model was released in parallel with a fine-tuned version of the instructions and a pre-trained checkpoint. While the model is not designed for complex dialog use cases, it has excellent base command adherence and responds to generic commands "out of the box".

4. Quantification of availability for production.The model provides Quantization-Aware Trained (QAT) checkpoints and supports running at INT4 accuracy with minimal performance loss, which is critical for deployments on resource-constrained devices (e.g., cell phones, edge devices).

In other words, if a user has a high-capacity, well-defined mission, needs to be smart about cost and needs to iterate and deploy quickly, or has a need to protect privacy, he's a good candidate for the Gemma 3 270M.

Hugging Face Address:
https://huggingface.co/collections/google/gemma-3-release-67c6c6f89c4f76621268bb6d

Lightweight models unleash end-side intelligence

此前，谷歌Gemma开源模型加速迭代：先是适用于单云和桌面加速器的Gemma 3和Gemma 3 QAT发布，随后是将强大实时multimodalAI直接引入边缘设备的Gemma 3n推出，本次Gemma 3 270M的推出填补了轻量模型版块。

Lightweight models are shattering parameter myths. While the perception of "parameter size determines performance" has long existed in the large model world, the Gemma 3 270M demonstrates the ability of small models to follow instructions and the power of fine-tuning. Starting with lightweight yet powerful models, users can build production systems that are lean, fast, and cost significantly less to run.

Google says the Gemma 3 270M can be found on Hugging Face, Docker, Kaggle, Ollama, and LM Studio, offering pre-trained and instruction-tuned versions for download.

Q&A

Q1: What's so special about the Gemma 3 270M?

A: Gemma 3 270M is an ultra-small AI model released by Google DeepMind that has only 270 million parameters but can run offline on low-power devices such as smartphones. It combines 170 million embedded parameters and 100 million Transformer block parameters to handle complex domain-specific tasks despite its small size.

Q2: How does the Gemma 3 270M perform?

A: In the IFEval benchmark, the Gemma 3 270M scored 51.2%, significantly outperforming other smaller models of the same size. Tests on the Pixel 9 Pro phone showed that 25 conversations consumed only 0.75% of battery power, demonstrating excellent energy efficiency.

Q3: How can developers use Gemma 3 270M?

A: Developers can access the model at Hugging Face, Docker, Kaggle, and other platforms that support rapid fine-tuning for specific tasks such as creative writing, sentiment analysis, entity extraction, and more. Google also provides full documentation, deployment guides, and support for various development tools.