Google releases Gemma 4, open-source on-device AI with agentic tool use for phones
Google released Gemma 4, an open-source multimodal model that runs entirely on smartphones without sending data to the cloud. The E2B and E4B variants require just 6GB and 8GB of RAM respectively and can autonomously use tools like Wikipedia, maps, and QR code generators through built-in agent skills. The model is available free via the Google AI Edge Gallery app for Android and iOS.
How Gemma 4 E2B ranks now
| Benchmark | Score | Rank | Source |
|---|---|---|---|
| GPQA | 43.4% | 157 of 162 | Hugging Face model card ↗ |
| MMLU-Pro | 60.0% | 70 of 73 | Hugging Face model card ↗ |
| LiveCodeBench | 44.0% | 45 of 48 | Hugging Face model card ↗ |
| AIME 2026 | 37.5% | 26 of 26 | llm-stats, Jul 2026 ↗ |
Rank is against every model in our database with a sourced score on that benchmark, each counted once with its best score.
Google releases Gemma 4, open-source on-device AI with agentic tool use for phones
Google released Gemma 4, an open-source multimodal AI model that processes text, images, and audio entirely on-device with autonomous tool-use capabilities. The model family includes four variants optimized for different hardware, with the smallest versions running on smartphones with as little as 6GB RAM.
Model specifications and variants
Gemma 4 ships in four sizes: E2B and E4B for smartphones, plus 26B and 31B models for servers. The "E" designates "effective parameters," referring to parameters active during inference rather than total parameter count.
The E2B variant consumes approximately 1.3GB quantized storage and runs on devices with 6GB RAM, while E4B requires roughly 2.5GB and 8GB RAM respectively. Google optimized both versions with Arm and Qualcomm for current mobile processors. According to Google, Gemma 4 on Android runs up to 4x faster than the previous generation while reducing battery consumption by up to 60 percent. Arm's benchmarks report even larger gains—an average 5.5x speedup on devices with newer Arm chips supporting the SME2 instruction set.
The 26B variant uses a mixture-of-experts architecture with 128 experts, keeping only 3.8 billion parameters active at inference time. The dense 31B model offers a 256,000-token context window.
All models process text, images, and audio across more than 140 languages. The entire Gemma family has achieved over 400 million downloads since initial release, according to Google.
On-device agentic capabilities
Gemma 4's defining feature is autonomous tool use without cloud connectivity. The bundled Google AI Edge Gallery app includes "agent skills"—built-in tools the model can independently invoke: Wikipedia search, interactive maps, auto-generated summaries, flashcard generation, and QR code generation. The model can describe photos, convert spoken input into diagrams and visualizations, and coordinate with other local models for text-to-speech or image generation.
The model automatically infers user intent and activates the appropriate skill. While tool invocation requires internet connectivity, the model itself runs entirely locally, and conversation history never persists.
Developers can create custom skills via GitHub and share them with the community. The app requires Android 12 or iOS 17 and has already reached fourth place among the most-downloaded free productivity apps in the iOS App Store, behind Claude, Gemini, and ChatGPT.
Google's demos show improvements in optical character recognition and time-aware reasoning, capabilities important for calendar, reminder, and alarm functionality.
Licensing and platform strategy
Gemma 4 releases under the commercially friendly Apache 2.0 license. Google built the models on research underlying its proprietary Gemini 3 system but made them freely available as open-source.
The E2B and E4B variants serve as the foundation for Gemini Nano 4, the next generation of Android's system-wide on-device model. Code written for Gemma 4 will work with Gemini Nano 4 upon release on flagship devices later in 2025. Gemini Nano currently runs on over 140 million Android devices, powering features like Smart Replies and audio summaries.
In December 2024, Google previewed a related approach with FunctionGemma, a 270-million-parameter model that translates natural language into structured function calls for phone tasks—toggling flashlights, creating contacts, managing calendars, and opening settings.
What this means
Gemma 4 marks a significant shift in on-device AI strategy. By combining unrestricted open-source licensing with genuine agentic capabilities at scale, Google enables developers to build privacy-preserving applications without cloud dependencies. The 4x speed improvements and 60 percent battery gains make the technology practical for mainstream phones. The model's integration pathway into Gemini Nano signals Google's commitment to making on-device AI standard across Android. For users, this means AI assistance that never transmits conversations to servers—a direct response to privacy concerns and a competitive move against cloud-dependent systems.
Related Articles
Google rolls out Gemini 4 Carbon to AI Ultra subscribers at $2/$10 per 1M tokens, Argon stays
Google is rolling out Gemini 4 Carbon to AI Ultra subscribers in the Gemini app, AI Mode and Antigravity, following last month's preview. Introductory API pricing is $2 per 1M input tokens and $10 per 1M output tokens until November 30, and Argon remains available for AI Pro and up.
Zyphra releases ZONOS3-Mini, an Apache 2.0 TTS model with 100M active parameters
Zyphra released ZONOS3-Mini, an Apache 2.0 text-to-speech model with 800M total parameters and 100M active parameters. It clones voices, steers delivery word by word and runs in real time in the browser.
MBZUAI's IFM releases PAN 2, an Apache 2.0 world model with a 29B diffusion transformer
The Institute of Foundation Models (IFM) at MBZUAI released PAN 2, an Apache 2.0 world model. It couples an LLM planner with a 29B diffusion transformer to simulate up to 20 seconds of action outcomes as video.
Fish Audio releases Live 3, a full-duplex voice model that reacts in as little as 186 ms
Fish Audio released Live 3, a full-duplex voice model for voice agents. It listens while it speaks, separates "mm-hmm" from real interruptions, and can react in as little as 186 ms. It is English only at launch.
Comments
Loading...