Google Releases Official Gemma 4 Models That Actually Fit on Your Phone
Squeezing a massive neural network into a smartphone used to require a PhD and a prayer. Now, Google is doing the dirty work themselves, promising fully functional, pocket-sized AI without turning your device into an expensive hand-warmer.
Google DeepMind uploaded the official quantized versions of its Gemma 4 family directly to Hugging Face, targeting laptops, edge devices, and mobile hardware. The most compact mobile variation squeezes down to a mere 1 GB of memory, meaning last year's budget Android phone might actually run it without instantly combusting.
While independent developers have been manually squishing models for years with varying degrees of success, this release marks the first time Google has optimized the compression themselves. The engineers utilized Quantization-Aware Training (QAT), a method where the AI is trained to handle low-precision calculations from day one, rather than trying to aggressively lobotomize a fully-grown model after the fact.
The newly released package includes five different sizes ranging from E2B up to 31B, offered in four specialized formats. This includes GGUF for quick local setups and mobile-targeted wNa8o8 formats designed to run smoothly on day-one supported engines like llama.cpp, Ollama, and LiteRT-LM.
Google claims this specialized training preserves model quality close to the uncompressed bfloat16 standard, though independent benchmarks have yet to verify if these mini-models are actually smart or just confidently incorrect.
Offline AI is finally moving from a niche hobbyist obsession to a mainstream corporate reality, leaving users to wonder whether a 1 GB brain in their pocket is a revolutionary tool for offline roaming or just another way to drain a smartphone battery in under twenty minutes.
Source: Hugging Face
Comments
This is where the magic happens: AI reads your discussion and rewrites the article based on the most interesting comments. Each strong comment adds points to the meter below. Once the meter is full, the article updates live — no page reload needed.