The open-source large model ecosystem has seen a major breakthrough in its underlying architecture. Google DeepMind recently officially released its most powerful open model to date, Gemma4. Although the parameter scale of this model remains similar to its predecessor, around 30 billion, its "intelligent density per parameter" has made a significant leap. Its performance on multiple core tasks is already comparable to top closed-source large models from one and a half years ago.

The most remarkable technological innovation of Gemma4 is the introduction of a new "E2B" (parameter unloading) architecture. In traditional Transformer architectures, the large embedding layer often consumes a massive amount of GPU memory. The new architecture cleverly adds an embedding table in each layer, replacing heavy full matrix multiplication calculations with a lookup table mechanism. For example, with a 50 billion parameter model, under the E2B architecture, only 20 billion parameters need to be loaded into GPU memory, while the remaining 30 billion can be safely unloaded to the CPU or even disk. This means the model can achieve fast inference with just 2GB of GPU memory, completely breaking through deployment bottlenecks on edge devices such as mobile phones, smartphones, and Raspberry Pi.

As a highly ambitious complex release, the Google DeepMind team coordinated nearly 50 external partners, including Hugging Face, llama.cpp, Ollama, NVIDIA, and AMD. Currently, Gemma4 has achieved deep integration with Android Studio. Developers can securely call AI to write Android code locally in offline environments without uploading any code to a cloud API in Agent mode, greatly meeting the rigid demand for data privacy and offline work in the workplace.

In terms of multimodal capabilities and core experiences, Gemma4 inherits the research achievements of Gemini3. Even small edge models with 2B or 4B parameters have excellent multilingual (supporting 140 languages) and multimodal understanding capabilities, capable of easily handling speech recognition, voice questions, and video analysis of 30 to 60 seconds. Although the model still lags behind large models in absolute knowledge capacity, and faces industry-recognized challenges in cutting-edge experimental explorations such as text diffusion (Diffusion Transformer) and expert mixture models (MoE) fine-tuning, its high-density intelligence is no longer negligible.

As the out-of-the-box capabilities of large models continue to improve, the vertical domain development ecosystem is undergoing a profound restructuring, and the heat of pure traditional fine-tuning is gradually cooling down. Looking ahead, Google DeepMind has made a milestone prediction: within the next 1 to 2 years, users' smartphones will be able to run powerful models equivalent to the performance of Gemini3Pro directly on the device. At that time, most complex intelligent agent tasks will be completed directly on the device without relying on cloud computing power, which will undoubtedly bring disruptive changes to the next generation of consumer application integration and user experience.