Google’s New AI Push Runs From Tiny Devices to Frontier Systems

Google is pushing its AI ambitions in two directions at once: toward powerful frontier systems and smaller models that can run directly on devices. The company has launched EmbeddingGemma 2 while announcing Gemini 4 Argon, creating a wide-ranging update across on-device search, coding, cybersecurity, and enterprise work.

The announcement arrives as Google tests how far its newest models can go in practical settings. EmbeddingGemma 2 targets useful tasks on devices, while Gemini 4 Argon enters a trusted-partner program focused on demanding work and the risks that come with frontier AI.

EmbeddingGemma 2 Brings Multimodal Search On Device

EmbeddingGemma 2 is Google’s next on-device AI model, with 740 million parameters. Its reach goes beyond text, covering coding, images, video, and audio, giving the model a broader set of materials to connect and search.

That broader design points toward a more natural way to find information stored on a device. A person could search for a video clip using a voice memo, or look through audio recordings with a text prompt. Instead of treating each format as a separate collection, the model is designed to connect these different types of content during an on-device task.

The model’s 740 million parameters also frame the challenge Google is tackling. EmbeddingGemma 2 must support multimodal work while operating on a device, rather than relying on a separate system for every search across coding, images, video, and audio.

This approach gives on-device AI a direct role in finding personal content. Voice, text, video, and audio can all become part of the same search experience, with the model handling the connections on the device.

Gemini 4 Argon Targets High-Stakes Work

Alongside EmbeddingGemma 2, Google announced Gemini 4 Argon, the first model in the 4 series. Google called the series its “next era of frontier intelligence,” and Argon is aimed at work where model performance can affect major technical and business decisions.

Google says Argon delivers frontier performance in coding, cybersecurity defense, and enterprise knowledge work in areas such as legal and finance. Employees inside Google are already using the model for debugging and large-scale codebase migrations, putting its capabilities into demanding software tasks.

Those internal uses have produced a concrete result. Argon helped engineers free up more than 300 tebibytes of memory across Google’s data centers without new hardware. That figure gives the model a role beyond generating code or answering questions: it has also supported work tied to the way Google operates its computing systems.

Google said Argon ties with OpenAI’s GPT-6 Astra for the highest score on the CWE-bench. The company also claims that Argon sets a new high score on a test measuring how models perform in real-world, long-horizon engineering tasks.

Performance claims like these place Argon in direct competition with other frontier models, but Google has not specified when Gemini 4 Argon will become publicly available. For now, the model is available to a set of trusted partners through Google’s Fairwind Program.

Safety Work Moves Alongside Model Testing

Google’s Argon announcement also focuses on how the company will assess and control the model before a broader launch. Google said it was “actively engaged” in the US government’s early-access framework for assessing cybersecurity and other risks tied to frontier models before launch.

Inside the model itself, Google implemented systems to track Argon’s chain of thought and stop it from performing when necessary. These controls aim to reduce misalignment risks as the model handles coding, cybersecurity defense, and other high-stakes work.

Google also said its safety measures include precautions around how Argon receives feedback. The goal is to avoid teaching the model to evade monitoring, a concern that becomes more important when a system is tested on complex tasks over long periods.

The company’s path to Argon also carries a missed release. Google had promised to release Gemini 3.5 Pro in June, but that release was postponed and Gemini 3.5 Pro was not released.

Now, Google is testing Argon with trusted partners while its employees use it for debugging, codebase migrations, and data-center work. At the same time, EmbeddingGemma 2 is bringing multimodal search to on-device tasks. Together, the two launches show Google pursuing AI that works across both ends of the computing spectrum.

The next milestone will be Argon’s public availability, but Google has not provided a date. Until then, the company’s latest AI push rests on two promises: EmbeddingGemma 2 can make device-based search more capable, while Gemini 4 Argon aims to raise the ceiling for frontier performance without leaving safety testing behind.

Stay connected via Google News

Follow us for the latest AI News