Google Ships Free On-Device AI Apps for iPhone via Gemma Models

Google releasing two AI-powered apps on the iOS App Store is not merely a product launch -- it is a strategic land grab inside Apple's ecosystem. While Apple and Google signed a deal in January 2026 for Gemini to power Apple Intelligence features, Google is simultaneously building a parallel AI stack that runs entirely outside Apple's control. AI Edge Gallery lets iPhone users download and interact with open-source Gemma models directly, bypassing Apple's own on-device intelligence layer entirely. This is Google establishing developer and user mindshare for its AI models on a competitor's hardware.

The move is especially significant because it arrives with zero marketing fanfare. There was no keynote, no press release, no blog post for Eloquent specifically. Google appears to be testing whether a genuinely superior free product can achieve organic adoption on iOS. If Eloquent gains traction as the default dictation tool for iPhone power users, Google will have inserted its AI models into millions of Apple devices -- creating a distribution channel for Gemma that Apple cannot easily shut down without appearing anti-competitive. The quiet launch also sidesteps the kind of scrutiny that a splashy announcement would invite from regulators already watching Big Tech AI moves closely.

Eloquent's business model -- or rather, its deliberate lack of one -- poses an existential threat to the subscription dictation market. Competing products like Wispr Flow and Willow charge approximately $15 per month, while SuperWhisper costs $85 per year. These apps rely on cloud processing, meaning they require internet connectivity and send voice recordings to external servers. Eloquent does everything these apps do, adds automatic filler word removal and text polishing, charges nothing, has no usage caps, and keeps all data on-device. The value proposition gap is not incremental; it is categorical.

This follows a familiar Google pattern: subsidize a high-quality free product to establish platform dominance, then monetize adjacent services. Eloquent includes an optional cloud mode powered by Gemini models for enhanced text polishing, which likely serves as the on-ramp to Google's broader paid AI ecosystem. For the dictation app market specifically, the implication is stark. Companies charging subscription fees for capabilities that a free, offline alternative now matches will need to either dramatically differentiate on features Google cannot easily replicate, or accept that the standalone dictation app category is being commoditized. The broader signal is that any AI application layer that merely wraps a model capability -- without unique data, workflow integration, or network effects -- is vulnerable to being zeroed out by foundation model providers releasing free reference apps.

While Eloquent grabbed headlines, the more consequential technical development may be FunctionGemma -- a 270-million-parameter model purpose-built for on-device function calling. Running at 1,916 tokens per second prefill and 142 tokens per second decode on a Pixel 7 Pro, this model can parse user intent and trigger real device actions without any server communication. Google demonstrated this through 'Mobile Actions' and 'Tiny Garden' features in AI Edge Gallery, but the implications extend far beyond demos.

On-device function calling solves two problems that have kept mobile AI agents theoretical. First, latency: cloud round-trips introduce delays that make real-time agentic interactions feel sluggish, whereas FunctionGemma responds locally in milliseconds. Second, reliability: an agent that depends on connectivity fails in elevators, airplanes, tunnels, and rural areas -- precisely the contexts where hands-free AI assistance is most valuable. As Google product manager Alice Zheng stated, shifting tool-use on-device allows developers to build interactions that respond instantly while remaining fully functional regardless of connectivity. Combined with Gemma 4 E2B's ability to run in under 1.5GB of memory, the hardware floor for running agentic AI is now a mid-range smartphone from 2022. This positions Google to define the on-device agent runtime before Apple, Meta, or anyone else establishes an alternative standard.

The Gemma model family has been downloaded over 400 million times with more than 100,000 community variants, but Gemma 4 represents a step change that makes consumer on-device deployment genuinely viable. The E2B model (2.3 billion effective parameters) runs in under 1.5GB of memory with 2-bit and 4-bit quantization, can process 4,000 tokens across two skills in under 3 seconds, and supports a 128K context window. The larger 31B dense model scores 85.2% on MMLU Pro -- a 26% improvement over Gemma 3's 67.6% -- and 89.2% on AIME 2026, placing it in territory previously reserved for models requiring data center hardware.

The collaboration with chip partners underscores that this is not just a model release but an infrastructure play. Qualcomm optimized Gemma 4 E2B and E4B for Dragonwing IQ8 NPU support, while MediaTek is working on mobile chipset optimization and NVIDIA is accelerating deployment through RTX AI Garage. All four Gemma 4 variants ship under Apache 2.0, meaning any device manufacturer or app developer can integrate them without licensing fees. The combination of dramatically reduced memory requirements, hardware-specific optimization from major chip vendors, and permissive licensing removes the three biggest barriers that previously kept generative AI tethered to the cloud. For developers, the practical takeaway is that building offline-first AI features is no longer a compromise -- it is now a competitive advantage.